U.S. patent application number 13/691313 was filed with the patent office on 2013-07-18 for interactive and automated tissue image analysis with global training database and variable-abstraction processing in cytological specimen classification and laser capture microdissection applications.
This patent application is currently assigned to LIFE TECHNOLOGIES CORPORATION. The applicant listed for this patent is Life Technologies Corporation. Invention is credited to David H. Kil.
Application Number | 20130182922 13/691313 |
Document ID | / |
Family ID | 31994137 |
Filed Date | 2013-07-18 |
United States Patent
Application |
20130182922 |
Kind Code |
A1 |
Kil; David H. |
July 18, 2013 |
Interactive and automated tissue image analysis with global
training database and variable-abstraction processing in
cytological specimen classification and laser capture
microdissection applications
Abstract
A system and method for performing tissue image analysis and
region of interest identification for further processing
applications such as laser capture microdissection is provided. The
invention provides three-stage processing with flexible state
transition that allows image recognition to be performed at an
appropriate level of abstraction. The three stages include
processing at one or more than one of the pixel, subimage and
object levels of processing. Also, the invention provides both an
interactive mode and a high-throughput batch mode which employs
training files generated automatically.
Inventors: |
Kil; David H.; (Prospect,
KY) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Life Technologies Corporation; |
Carlsbad |
CA |
US |
|
|
Assignee: |
LIFE TECHNOLOGIES
CORPORATION
Carlsbad
CA
|
Family ID: |
31994137 |
Appl. No.: |
13/691313 |
Filed: |
November 30, 2012 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
10662765 |
Sep 15, 2003 |
8346483 |
|
|
13691313 |
|
|
|
|
60410433 |
Sep 13, 2002 |
|
|
|
Current U.S.
Class: |
382/128 |
Current CPC
Class: |
G01N 15/1475 20130101;
G01N 2001/284 20130101; G06T 2207/30024 20130101; G06T 7/0012
20130101; G01N 1/06 20130101; G06T 2207/10056 20130101; G06K
9/00147 20130101; G01N 1/30 20130101; G06T 2207/20132 20130101;
G06K 9/0014 20130101; G06K 9/00127 20130101; G06T 7/143 20170101;
G02B 21/365 20130101; G06K 9/6253 20130101; G06T 7/00 20130101;
G06T 7/11 20170101; G06K 9/6228 20130101 |
Class at
Publication: |
382/128 |
International
Class: |
G06K 9/00 20060101
G06K009/00 |
Claims
1-18. (canceled)
19. A computer-implemented method for image analysis, the
computer-implemented method comprising: receiving a first image;
transforming the first image into a feature space; selecting a
level of abstraction; selecting a database containing parameters
based on the selected level of abstraction; classifying the first
image into regions of interest employing the parameters from the
database based on the selected level of abstraction; updating the
parameters of the database for the level of abstraction with data
from the first image; receiving a second image; transforming the
second image into a feature space; classifying the second image
into regions of interest employing the updated parameters from the
database based on the selected level of abstraction; updating the
parameters of the database with data from the second image.
20. The computer-implemented method for image analysis of claim 19
wherein selecting the level of abstraction includes selecting pixel
processing.
21. The computer-implemented method for image analysis of claim 20
further including transmitting the regions of interest obtained
from pixel processing for laser capture microdissection.
22. The computer-implemented method for image analysis of claim 19
wherein selecting the level of abstraction includes selecting
subimage processing.
23. The computer-implemented method for image analysis of claim 22
wherein classifying the first image includes classifying the first
image into regions of interest employing parameters from the
database for pixel processing and classifying the first image into
regions of interest employing parameters from the database for
subimage processing; and wherein classifying the second image
includes classifying the second image into regions of interest
employing parameters from the database for pixel processing and
classifying the second image into regions of interest employing
parameters from the database for subimage processing.
24. The computer-implemented method for image analysis of claim 23
further including transmitting the regions of interest obtained
from subimage processing for laser capture microdissection.
25. The computer-implemented method for image analysis of claim 19
wherein selecting the level of abstraction includes selecting
object processing.
26. The computer-implemented method for image analysis of claim 25
wherein classifying the first image includes classifying the first
image into regions of interest employing parameters from the
database for pixel processing and classifying the first image into
regions of interest employing parameters from the database for
subimage processing and classifying the first image into regions of
interest employing parameters from the database for object
processing; and wherein classifying, the second image includes
classifying the second image into regions of interest employing
parameters from the database for pixel processing and classifying
the second image into regions of interest employing parameters from
the database for subimage processing and classifying the second
image into regions of interest employing parameters from the
database for object processing.
27. The computer-implemented method for image analysis of claim 26
further including transmitting the regions of interest obtained
from object processing for laser capture microdissection.
28. The computer-implemented method for image analysis of claim 25
wherein classifying the first image includes classifying the first
image into regions of interest employing parameters from the
database for pixel processing and classifying the first image into
regions of interest employing parameters from the database for
object processing; and wherein classifying the second image
includes classifying the second image into regions of interest
employing parameters from the database for pixel processing and
classifying the second image into regions of interest employing
parameters from the database for object processing.
29. The computer-implemented method for image analysis of claim 28
further including transmitting the regions of interest obtained
from object processing for laser capture microdissection.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application claims priority to U.S. Provisional
Application Ser. No. 60/410,433, entitled "INTERACTIVE AND
AUTOMATED TISSUE IMAGE ANALYSIS WITH GLOBAL TRAINING DATABASE AND
VARIABLE-ABSTRACTION PROCESSING IN CYTOLOGICAL SPECIMEN
CLASSIFICATION AND LASER CAPTURE MICRODISSECTION APPLICATIONS",
filed on Sep. 13, 2002 which is incorporated herein by reference in
its entirety.
FIELD OF THE INVENTION
[0002] The invention relates generally to automated tissue image
analysis, and in particular, to image analysis for cell
classification and laser capture microdissection applications.
BACKGROUND OF THE INVENTION
[0003] Laser capture microdissection (LCM) is a robust and reliable
technology for isolating pure populations of cells from
heterogeneous tissue samples for subsequent analysis. The LCM
technology integrates a laboratory microscope with a low-energy
laser and transfer film in a convenient one-step, aim-and-shoot
method. Generally, thin sections of tissue samples are mounted on
standard glass slides employing various common methods known in the
art such as fixing tissues with alcohol-based precipitation
techniques. LCM is compatible with various common methods for the
preparation of tissue sections.
[0004] The thin tissue sections may be stained by standard
techniques such as hematoxylin and eosin, methylene green nuclear
stain, fluorescence in situ hybridization, or immunohistochemistry
for identification of tissue morphology and cell populations of
interest. Staining the sample may or may not be required. In some
cases, a marker is added to the tissue sample to adhere to a
specific type of site in the tissue to render the site detectable
in an image of the tissue that is captured via an acquisition
system. Markers may be antibodies, drugs, or other compounds that
attach or bind to the tissue component of interest and are
radioactive or fluorescent or have a distinctive color or otherwise
detectable. Once mounted on a substrate surface such as a standard
glass slide, the transfer film is located in juxtaposition to the
tissue surface. The transfer film is typically made of
thermoplastic film such as ethylene-vinyl acetate. Broadband energy
absorbing transfer films are described in U.S. Pat. No. 6,495,195
entitled "Broadband absorbing film for laser capture
microdissection" issued to Baer et al. and hereby incorporated by
reference in its entirety.
[0005] Once loaded in a LCM device, a tissue sample is viewed via a
microscope and a cell or cells of interest are targeted. The laser
is directed at the cell or cells of interest and pulsed to provide
enough energy to transiently and locally melt the thermoplastic
film and activate the transfer film in the precise focal region of
the laser beam. The laser beam spot size can be adjusted so that a
targeted individual cell or cluster of cells can be selected in one
or more pulses of the laser. The optical system of a LCM instrument
is described in U.S. Pat. No. 6,215,550 and U.S. Pat. No. 6,512,576
both entitled "Laser capture microdissection optical system" and
issued to Baer et al. and both hereby incorporated by reference in
its entirety.
[0006] The cell or cells in the activated region of the transfer
film adhere to the transfer film and can be extracted from the
remaining tissue sample with the unselected tissue remaining in
contact with the glass slide. Because the thermoplastic film
absorbs most of the thermal energy and the pulse lasts for a
fraction of a second, no detectable damage of the biological
macromolecules occurs. Once removed from the tissue sample, the
selected cell or cells are subjected to appropriate extraction
conditions for ensuing molecular analysis. To facility subsequent
extraction steps, the transfer film can be mounted on a substrate
surface that is shaped like a cap that fits a microcentrifuge tube
as described in U.S. Pat. No. 6,157,446 entitled "Laser capture
microdissection analysis vessel" issued to Baer et al. and hereby
incorporated by reference in its entirety and in U.S. Pat. No.
5,859,699 entitled "Laser capture microdissection analysis vessel"
issued to Baer et al. and hereby incorporated by reference in its
entirety. A method for manufacturing a consumable is described in
U.S. Pat. No. 5,985,085 entitled "Method of manufacturing
consumable for laser capture microdissection" issued to Baer et al.
and hereby incorporated by reference in its entirety. Laser capture
microdissection is also described in U.S. Pat. No. 6,469,779
entitled "Laser capture microdissection method and apparatus"
issued to Baer et al. and hereby incorporated by reference in its
entirety.
[0007] By isolated only target cells from the tissue sample using
LCM, researchers can immediately analyze the gene and enzyme
activity of the target cells using other research tools. Such
procedures as polymerase chain reaction amplification of DNA and
RNA, and enzyme recovery from the tissue sample have been
demonstrated. No limitations have been reported in the ability to
amplify DNA or RNA from tumor cells extracted with laser capture
microdissection. LCM has been particularly advantageous in
identifying the differences between expression levels in normal and
diseased tissues. In addition to combining LCM with several genomic
and proteomic techniques to document the progression of normal
cells to premalignant and metastatic cancer cells in various
tissues, microdissected cells are also used in applications for
gaining new insights in developmental biology.
[0008] The LCM technique has been automated as described in
International Patent Publication No. WO 01/33190 entitled
"Automated laser capture microdissection" to Baer et al. and hereby
incorporated by reference in its entirety and in International
Patent Publication No. WO 02/037159 entitled "Road map image for
automated microdissection" to Baer et al. and hereby incorporated
by reference in its entirety. Continued automation of the LCM
process is desired. In particular, automated tissue image analysis
for target cell classification for subsequent laser capture
microdissection is wanting for high-throughput batch processing.
This invention addresses these needs for increased automation,
accurate and reliable image analysis and cell classification for
LCM.
[0009] Tissue analysis and identification of a cell or a region of
interest (ROI) have always been a time-consuming, laborious
process. The major obstacles to the successful deployment of a
high-throughput tissue analysis system are the diversity in the
ROIs and cell types, the variability in staining, and the
skepticism from the user community.
[0010] There exists an abundance of literature and prior art in the
field of automated tissue recognition. U.S. Pat. No. 6,327,377
issued to Rutenberg, et al. entitled "Automated cytological
specimen classification system and method" uses a primary detector
based on thresholding of an integrated optical density (IOD), a
secondary classifier that utilizes a three-layer back-propagation
neural network for pattern matching, and a tertiary screener by a
human operator. Another U.S. Pat. No. 6,215,892 issued to Douglas
et al. entitled "Method and apparatus for automated image analysis
of biological specimens" uses a color-ratio threshold as an initial
detector followed by a morphology-based analysis for identifying
potential ROI candidates. U.S. Pat. No. 5,987,158 issued to Meyer
et al. entitled "Apparatus for automated identification of thick
cell groupings on a biological specimen" takes a slightly different
approach to ROI classification. After image segmentation, it uses
Fisher's linear binary decision tree in series to perform object
(ROI) classification.
[0011] Unfortunately, none of these patents address the core issue
of how to facilitate high-throughput cell classification processing
that yields robust performance through the creation and
manipulation of global training databases to ease the burden on
human operators. Furthermore, modern nonparametric learning
algorithms for studying gene activation patterns and regulatory
networks require a lot of high-quality data. This invention bridges
the gap by providing a flexible, high-throughput cell
classification processing chain in two complementary dimensions to
improve system performance with age.
BRIEF SUMMARY OF THE INVENTION
[0012] In accordance with one aspect of the invention, there is
provided a computer method for image analysis. The method includes
the step of receiving an image. The image is transformed into a
feature space. At least one region of interest (ROI) at a pixel
level of processing is selected and features from the ROI at a
pixel level of processing are extracted. Also, at least one non-ROI
at a pixel level of processing is selected and features from the
non-ROI at a pixel level of processing are extracted. The extracted
features are ranked based on feature performance for successful
detection of a selected ROI at a pixel level of processing and the
ranked extracted features are recorded. A classification algorithm
is selected and executed to classify the image into regions of
interest at a pixel level of processing. The resulting ROIs based
on pixel processing are recorded for further processing.
[0013] In accordance with another aspect of the invention, there is
provided a computer method for image analysis. The method includes
the step of receiving a first image. The first image is transformed
into a feature space and a level of abstraction is selected. A
database containing parameters based on the selected level of
abstraction is also selected. The first image is classified into
regions of interest employing the parameters from the database
based on the selected level of abstraction. The parameters of the
database for the level of abstraction are updated with data from
the first image. The method further includes the step of receiving
a second image and transforming the second image into a feature
space. The second image is classified into regions of interest
employing the updated parameters from the database based on the
selected level of abstraction. The parameters of the database are
updated with data from the second image.
BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS
[0014] The foregoing and other advantages of the invention will
become apparent upon reading the following detailed description and
upon reference to the drawings in which:
[0015] FIG. 1 is a schematic diagram of the apparatus of the
invention;
[0016] FIG. 2 is a perspective view of a portion of the automated
LCM device;
[0017] FIG. 3 is a exploded top level block diagram of a portion of
the automated LCM device;
[0018] FIG. 4 is a block diagram representation of a method of
classifying objects according to the invention;
[0019] FIG. 5 is a block diagram representation of a method of
pixel processing according to the invention;
[0020] FIG. 6 is a block diagram representation of a method of
run-classification under the "PIXEL LEARN" processing option
according to the invention;
[0021] FIG. 7a is a depiction of an example of a multi-modal
feature distribution;
[0022] FIG. 7b is a depiction of an example of a unimodal feature
distribution;
[0023] FIG. 8a is one example of a visual output of pixel
processing according to the invention;
[0024] FIG. 8b is one example of a visual output of pixel
processing that includes region labeling using color according to
the invention;
[0025] FIG. 9 is a block diagram representation of one example of a
method for database management according to the invention;
[0026] FIG. 10 is a block diagram representation of one method of
run-classification under the "PIXEL CLASSIFY" processing option
according to the invention;
[0027] FIG. 11 is a block diagram representation of a method of
subimage processing according to the invention;
[0028] FIG. 12 is an one example of a visual output during subimage
processing that includes polygonal ROI and non-ROI segments
according to the invention;
[0029] FIG. 13 is an is a block diagram representation of a method
of run-classification under the "SUBIMAGE LEARN" processing option
according to the invention;
[0030] FIG. 14a is one example of a visual output of subimage
processing according to the invention;
[0031] FIG. 14b is one example of a visual output of pixel
processing according to the invention;
[0032] FIG. 15 is a block diagram representation of one method of
run-classification under the "SUBIMAGE CLASSIFY" processing option
according to the invention;
[0033] FIG. 16 is a block diagram representation of a method of
object processing according to the invention;
[0034] FIG. 17 is a block diagram representation of one method of
run-classification under the "OBJECT LEARN" processing option
according to the invention;
[0035] FIG. 18 is one example of a visual output of object
processing according to the invention;
[0036] FIG. 19 is a block diagram representation of one method of
run-classification under the "OBJECT CLASSIFY" processing option
according to the invention; and
[0037] FIG. 20 is one example of a visual output of the image, the
image after pixel processing, the image after subimage processing
and the image after object processing.
[0038] While the present invention is susceptible to various
modifications and alternate forms, specific variations have been
shown by way of example in the drawings and will be described
herein. However, it should be understood that the invention is not
limited to the particular forms disclosed. Rather, the invention is
to cover all modifications, equivalents, and alternatives falling
within the spirit and scope of the invention as defined by the
appended claims.
DETAILED DESCRIPTION OF THE INVENTION
[0039] This application claims priority to U.S. Provisional
Application Ser. No. 60/410,433, entitled "INTERACTIVE AND
AUTOMATED TISSUE IMAGE ANALYSIS WITH GLOBAL TRAINING DATABASE AND
VARIABLE-ABSTRACTION PROCESSING IN CYTOLOGICAL SPECIMEN
CLASSIFICATION AND LASER CAPTURE MICRODISSECTION APPLICATIONS",
filed on Sep. 13, 2002 which is incorporated herein by reference in
its entirety. This application also claims priority to U.S.
application Ser. No. 09/707,313 entitled "AUTOMATED LASER CAPTURE
MICRODISSECTION", filed on Nov. 6, 2000, which is incorporated
herein by reference in its entirety.
[0040] A schematic diagram of the apparatus 100 of the invention is
shown in FIG. 1. The apparatus 100 of the invention comprises a
laser capture microdissection (LCM) device 102 coupled to a central
processor 104. The processor 104 is connected to a variety of input
and output devices including a monitor 106, keyboard 108, a
pointing device 110 such as a mouse, and a printer 112. A hard disk
114 is connected and controlled by the processor 104. The processor
104 is an IBM PC/AT or compatible although it may be another
computer-type device suitable for efficient execution of the
functions described herein. The processor 104 controls the
functioning and flow of data between components of the LCM device
102, causes execution of a variety of classification and other
algorithms and handles storage of image and classification
information. The invention may be embodied in software, on a
computer readable medium or on a network signal, to be run on a
general purpose computer or on a network of general purpose
computers. As is known in the art, the neural network component may
be implemented with dedicated circuits rather than with one or more
general purpose computers.
[0041] With reference now to FIGS. 2 and 3, the LCM device 102,
which is connected to and controlled by the processor 104, will now
be discussed. The LCM device 200 comprises an illumination and
laser optical subsystem 202, 302 interconnected with a translation
stage 204, 304, a transfer film carrier supply 306, and a transfer
film carrier handling subsystem 206, 308.
[0042] The translation stage 204, 304 includes a work surface 310
adapted to receive at least one sample carrier 312. The sample
carriers 312 are depicted in FIG. 3 as standard microscope slides
having samples 314 attached thereto. The translation stage 304
defines at least one beam path hole (not shown) corresponding to an
at least one sample carrier receiving location 316 such that light
may pass through the sample 314, sample carrier 312 and translation
stage 304. A vacuum chuck, having one or manifold holes and
conduits at the carrier receiving location 316, is employed to
secure the sample carrier 312 in position when the vacuum is
engaged. The translation stage 304 is automated by a fore-and-aft
motor 318 and a translation stage side-to-side motor 320 both of
which are controlled by the central processor 104.
[0043] The transfer film carrier supply 306 is located adjacent to
the translation stage 304. The transfer film carrier supply 306 is
connected to the translation stage 304 or located on a nearby
element separate from the translation stage 304. The transfer film
carrier supply 306 is adapted to receive and serve as a staging
area for transfer film carriers 322. In FIG. 3, the transfer film
carriers 322 are depicted in the form of discrete consumables and
shaped in the form of caps that are adapted to mate with reaction
vessels for post-capture processing. The caps include a laser
capture microdissection transfer film (not shown) attached to the
lower surface of the cap.
[0044] The transfer film carrier handling subsystem 208, 308 is an
automated assembly for picking transfer film carriers 322 from the
transfer film carrier supply 306 and placing them in juxtaposition
to the tissue samples 314 located on sample carriers 312 prior to
cell capture. After cell capture, the transfer film carrier
handling subsystem 208, 308 removes the transfer film carriers 322
and delivers them to an output station 324 that is also located
adjacent to the translation stage 204, 308. The output station 324
is a standard microscope slide onto which the transfer film
carriers 322 are placed. Alternatively, the output station 324 is a
capping station for automatically mating the caps with reaction
vessels such as microcentrifuge tubes for post-LCM processing.
[0045] The transfer film carrier handling subsystem 208, 308 is
automated and controlled by the processor 104. The transfer film
carrier handling subsystem 208, 308 includes an arm 326 having a
working end adapted to pick and place a transfer film carrier 322.
The arm 326 is controlled by a carrier translation motor 328 and a
carrier lift motor 330 included in the subsystem 308. The arm 326
includes at least a load position and at least an inspect position.
While in the load position, the working end of the arm 326 is
located adjacent to the transfer film carrier supply 306. In this
position, the arm 326 is manipulated by the motors 328 and 330 to
pick up a transfer film carrier 322 from the transfer film carrier
supply 306. From the load position, the arm 326 is manipulated
automatically to the inspect position. While in the inspect
position, the working end of the arm 326 is located adjacent to the
sample carrier 312 and the arm 326 is manipulated to place the
transfer film carrier 322 into precise juxtaposition with the
sample 314. After LCM is complete, the arm 326 removes the transfer
film carrier 322 and is automatically manipulated to an unload
position. In the unload position, the working end of the arm 326 is
located adjacent to the output station 324 where it is manipulated
to release the transfer film carrier 322 and/or to mate the
transfer film carrier 322 with a reaction vessel. The sample
carrier handling subsystem 308 is described in greater detail in
U.S. patent application Ser. No. 09/707,313 filed on Nov. 6, 2000,
entitled "Automated Laser Capture Microdissection," the contents of
which is hereby incorporated by reference in its entirety.
[0046] The illumination and laser optical subsystem 202, 302, which
is connected to and controlled by the processor 104, comprises an
optical system 332 interconnected to a laser 334, an illuminator
336 and an image acquisition system 338. The illumination and laser
optical subsystem 302 is a modified inverted microscope and laser
system. The laser 334, for example, has laser diodes made of
gallium arsenide with aluminum doping (AlGaAs) and emits radiation
at approximately 800 nm. The laser 334 is controlled by the
processor 104 and is selectively activable to emit a laser beam
340.
[0047] The illumination and laser optical subsystem 302 also
includes an illuminator 336. The illuminator 336 is controlled by
the processor 104 and is selectively activable to emit light 342 to
illuminate the sample. In typical microscopes, the illuminator 336
is, for example, a condenser or tungsten-halogen lamp. According to
one aspect of the invention, the illuminator 336 is a light
emitting diode (LED). The illuminator 60 is particularly important
because of the need for consistency in the light source used to
illuminate samples especially when performing automated cell
recognition based on image processing using sophisticated software
employing color-based algorithms. Typically, the light source for a
microscope system is a halogen lamp. These lamps are low cost and
provide illumination at a color temperature of approximately 3200K
when operating at their designed voltage and power levels such as 6
volts and 30 watts. A limitation of these lamps is that as the
voltage and power levels are decreased, the color temperature of
their output shifts towards the red-end of the spectrum. This shift
in color causes the image of the sample illuminated in the
instrument to shift in color. The human eye can deal with this
color shift to some extent but an imaging system employing a camera
and image processing software cannot adapt so easily. Because the
camera is highly light-sensitive, the user will often reduce the
illuminator power in order to prevent saturation of the camera by
the illuminator. This reduction in power shifts the illuminator
color temperature. The camera system will display the colors based
on its baseline calibration, and if the illumination source is
red-shifted, then the camera image will also be red-shifted. The
shift in color can cause serious problems with image recognition
systems, particularly those employing learning files that are
applied across multiple images as will be discussed in detail
hereinbelow. For example, if the image recognition system learning
file was trained at one color temperature, then its ability to
recognize subsequent images taken at another color temperature
resulting from slightly different lamp power settings, can be
seriously impaired. In one variation of this invention, a white LED
is employed as the illuminator 336 because the spectral
characteristics of the LED do not change significantly with applied
current. The variation in color temperature with lamp intensity is
eliminated allowing the laser microdissection process including
image recognition to be enhanced. In another variation, multiple
LEDs of different colors, such as red, green and blue, are employed
and mixed together and used to continuously adjust the color of the
system by varying the amount of each color in the mix. In yet
another example, rotating crossed polarizers or variable neutral
density filters are located between the illuminator 336 and a
partially-transmissive and partially reflective surface 344 to
adjust the light intensity at a sample 314 while leaving the
illumination lamp 336 at its maximum setting.
[0048] The optical system 332 of the illumination and laser optical
subsystem 302 has an optical axis 346 as shown by the dashed line
in FIG. 3. The optical system 332 comprises at least a partially
transmissive and a partially reflective surface 344, a focusing
lens 348, and an objective 350. The partially transmissive and
partially reflective surface 344 is optically coupled to a focusing
lens 348 that is in turn optically coupled to an objective 350. The
optical system 332, in turn is optically connected to the laser
334, illuminator 336, and acquisition system 338.
[0049] The optical system 332, laser 334 (shown with reference
numeral 234 in FIG. 2), illuminator 336 and acquisition system 338
are configured such that the laser light beam 340 passes from the
laser 334 to the partially transmissive and partially reflective
surface 344. The partially transmissive and partially reflective
surface 344 is, for example, a dichroic mirror. From the partially
transmissive and partially reflective surface 344, the laser beam
340 passes through the focusing lens 348.
[0050] The focusing lens 348 is connected to a laser focus motor
352 which is controlled by the processor 104 and operates to
control the focusing lens 348 to focus and adjust the laser beam
spot size from a smaller tightly focused condition to a relatively
larger beam spot size arising from a relatively defocused laser
beam. In one variation, the focusing lens 348 is replaced with a
stepped lens. In another variation, a variable aperture is
optically coupled to the focusing lens 348 to adjust the laser beam
spot size.
[0051] From the focusing lens 348, the laser beam 340 is directed
along the optical axis 346 toward the translation stage 304. In
particular, the laser is precisely directed at targeted cells of a
tissue sample on a sample carrier 314 positioned in a sample
carrier receiving location 316. The laser beam 340 passes through
the beam path hole in the translation stage to a cut-off filter
that reflects or absorbs energy from the laser beam.
[0052] Light 342 from the illuminator 336 passes to the partially
transmissive and partially reflective surface 344 and then through
the focusing lens 348 along the optical axis 346 and through the
beam path hole of the translation stage 304. From the translation
stage 304, illumination light 342 passes through the objective 350
and then to the acquisition system 338. Light 342 from the
illuminator 336 and the laser light beam light 340 are superimposed
at the partially transmissive and partially reflective surface 344
along the optical axis 346. The objective 350 comprises an
objective changer with a series of microscope objectives deployed
on an objective turret wheel. The objective 350 is automated by an
objective changer motor 354 and controlled by the processor 104 to
select an appropriate objective lens. The objective 350 is also
automated by an objective focus motor 356 to focus illumination
light 342 passing through the beam path hole of the translation
stage 304.
[0053] In one variation, the optical system 332 further includes a
collimator lens and condenser lens (not shown) such that the
partially transmissive and partially reflective surface 344 is
located between the collimator lens and the condenser lens. Also, a
scattering media is included and located between the condenser lens
and the translation stage 304 and above the sample 314. As shown in
FIG. 3, the scattering media is a piece of diffuser glass 358, for
example, located integrally with the working arm 326 of the
transfer film carrier handling subsystem 308 and adapted to be
positioned in and out from within the optical axis 346.
Illumination light 342 is diffused by the scattering media
illuminating the sample from all angles. This high illumination
angle or high numerical aperture illumination provides high image
quality and eliminates the need for refractive index matching of a
sample. Such a scattering media allows visualization of the cell
nucleus and other sub cellular structures that would normally be
obscured by normal illumination techniques. The scattering media
is, for example, a diffuser material such as milk or opal glass,
frosted glass, standard printer/photocopier paper, a lenticular
sheet, a volume diffuser or a surface diffuser. In one variation,
the scattering media is integral with the transfer film carrier 322
or transfer film itself. Illumination light and laser beam light
steering mirrors 360 are employed where necessary to direct
light.
[0054] In one variation, the illumination and laser optical
subsystem 302 includes a fluorescent light source 362 coupled to
the optical system 332 and controlled by the processor 104. The
light source 362 is, for example, an EPI-fluorescent xenon or
mercury lamp that emits light 364 having a specific wavelength or
within a specific wavelength range. The specific wavelength or
wavelength range of a beam emitted by the light source 362 is
selected by a fluorescence filter wheel 366 operated by a
fluorescence filter changer motor 368 to excite a fluorescent
system (e.g., chemical markers and optical filtering techniques
that are known in the industry) that is incorporated in or applied
to the sample 314. The sample 314 includes at least one member
selected from the group consisting of chromophores and fluorescent
dyes, synthetic or organic and the frequency of the beam emitted by
the fluorescence laser 362 (shown as element 262 in FIG. 2) is
tuned to match the at least one member. The fluorescent laser beam
364 is superimposed with the laser beam 340 and light 342 from the
illuminator 336. Fluorescence emitted by the sample is optionally
amplified by the objective 350, reflected by a camera changer
mirror 360 and captured for viewing by the acquisition system 338.
The objective changer motor 354 and the objective focus motor 356
operate to adjust the fluorescent laser beam 364 and the emitted
fluorescent beam.
[0055] The acquisition system 338 includes one or more CCD cameras
positioned in the optical path to capture a focused, magnified
electronic image of an area of the sample. As shown in FIG. 3, the
acquisition system 338 includes a black-and-white camera 372, a
color camera 374 and a roadmap camera 376 for capturing appropriate
images. The acquisition system 338 is coupled to and controlled by
the processor 104 delivering images thereto for image processing,
viewing and navigation purposes. The color camera 374 may be a
conventional RGB camera, or other camera able to provide suitable
information of the specimen or image to the computer. The roadmap
camera 376 and its navigational functions and features are
explained in detail in U.S. patent application Ser. No. 09/707,313
filed on Nov. 6, 2000, entitled "Automated Laser Capture
Microdissection" and incorporated herein by reference in its
entirety. The LCM device includes an electronics panel (marked with
reference numeral 201 in FIG. 2) comprising printed circuit boards
and instructions for the automated LCM device 102, computer
interface cards and input/output devices for connecting and
communicating with the central processor 104.
[0056] The apparatus 100 operates to carry out the following
general steps for laser capture microdissection. First, a tissue or
sample smear 314 is fixed onto a standard microscope slide or
sample carrier 316 by routine protocols. Typically, the sample is
fluorescently stained. Histochemical staining is especially useful
for identification of cells of interest. Immunological labeling is
another method that is used to identify a cell of interest.
According to this technique, an antibody specific for an antigen
whose presence or absence is characteristic of a rare cell of
interest is bound to the cell and directly or indirectly labeled
with a fluorescent stain. Immunolabeling and staining techniques
are well known in the art and incorporated herein by reference.
[0057] Next, the sample carrier 314 is introduced into the LCM
device 102 at a sample carrier receiving location 316. A transfer
film is placed in juxtaposition to the sample, either in contact
with the sample or separated from the sample by a small distance.
Typically, the transfer film is attached to a substrate surface of
the transfer film carrier or cap. The transfer film is, for
example, an approximately 100-micron thick ethyl vinyl acetate
(EVA) film available from the Electroseal Corporation of Pompton
Lakes, N.J. The transfer film is manufactured containing organic
dyes that are chosen to selectively absorb in the near infrared
region of the spectrum overlapping the emission region of laser
that is employed, typically an AlGaAs laser. Basically, the
transfer film is selected and/or dyed or filled with
energy-absorbing matter to absorb energy at the wavelength emitted
by the laser.
[0058] The illuminator 336 illuminates the sample. Simultaneously,
the EPI-fluorescent lamp 362 is turned on to activate the
fluorescent markers in the sample 314. Also, the laser beam is
activated in idle mode such that the laser beam light provides a
visible low amplitude signal that can be detected during live
viewing via the acquisition system for visual alignment and
targeting or when adjustment of the beam spot size is desired. As
previously mentioned, the laser beam spot size is adjusted by
operating the laser focus motor 352 to move the focusing lens 348.
An image is capture via the acquisition system 338 and delivered to
the processor 104 for target cell detection. Target cell detection
is performed manually or automatically. Novel automated and
semi-automated target cell detection methods will be discussed in
detail hereinbelow. After target cells are detected and located,
the processor automatically directs the laser beam path such that
it is coincident with the location of target cells and
simultaneously activates the laser into pulse mode. In pulse mode,
the laser beam light activates the transfer film such that the
activated portion of the transfer film increases in temperature,
expands and adheres to the tissue region of targeted cells. Once
the desired target cell or cells are adhered to the transfer film,
the transfer film is lifted from the sample and the selected
portion of tissue comprising targeted cells is removed with the
film. The targeted cells are then ready for post-LCM
processing.
[0059] Automated and semi-automated target cell recognition methods
will now be discussed. FIG. 4 is a block diagram representation 400
of a method of classifying objects according to the invention in
which focused and magnified electronic images captured by one or
more cameras of the acquisition system are transmitted to the
processor for image processing and classification of the cells
appearing in the imaged area. The raw image, preferably represented
by electrical signals, is delivered to the computer processor 402.
The raw image data is stored in and retrieved from various
databases 414 as shown by line 401. The raw image data undergoes
image transformation 404 and is stored in one or more databases as
shown by line 403 before commencing classification under one or
more than one of the classification protocols: pixel processing
406, subimage processing 408, and object processing 410. One or
more classification protocols are executed for image processing and
classification of the cells appearing in the imaged area prior to
commencing the steps of a desired application 412 such as
extracting classified cells using laser capture microdissection.
Throughout the method of operation, various databases 414 are
generated and employed interactively along each processing stage,
pixel 406, subimage 408 and object 410 along lines 405, 407 and
409, respectively.
[0060] The invention performs region-of-interest (ROI)
identification in two processing dimensions. The first processing
dimension utilizes variable abstraction processing. Pixel
processing 406 is a first level of abstraction that relies on a
number of attributes, such as red (R), green (G), blue (B), hue
(H), saturation (S), intensity or value (V), gray intensity after
RGB-to-gray conversion and histogram equalization (GI), red/green
(RG), red/blue (RB), green/blue (GB), to perform classification.
Instead of using a single attribute, pixel processing uses an
appropriate composition of pixel attributes that are optimized
automatically for each database. Subimage processing 408 is a
second level of abstraction that utilizes attributes, such as
texture-related information and neighborhood statistics at an
arbitrary subimage level that is selected by the user that is
tailored to the tissue types in study, to exploit the level of
information available at the variable-size image level. Object
processing is a third level of abstraction that extracts features
at the object level such as morphology, gradient, and texture
information at the object level to perform ROI identification.
[0061] As shown by the arrows in FIG. 4, the transition from pixel
processing 406 to subimage processing 408 to object processing 410
is not a rigid transition. Instead, the invention employs a
flexible state transition so that processing steps are tailored to
the complexity of the image being analyzed. For example, pixel
processing 406 may be more than adequate for the classification of
a particular image dataset thereby eliminating the need for
subimage 408 or object processing 410. In such a case, processing
proceeds directly from pixel processing 406 to LCM 412 as shown by
arrow 416. If the classification performance results are inadequate
for a particular image data set, the processing may proceed from
pixel processing 406 to subimage processing 408 along line 418 and
then from subimage processing 408 to LCM along line 420.
Alternatively, image recognition processing may proceed from pixel
processing 406 to object processing 410 via arrow 419 and then from
object processing 410 to LCM 412 via line 424. Alternatively,
processing steps may proceed through all three levels of
abstraction before LCM 412 following arrows 418, 422 and 424.
[0062] A focused and magnified electronic image from the microscope
and captured by a camera of the acquisition system 338 is
transmitted to and digitized into a pixel data image by the
processor at block 402 and is stored in a database 414. The raw
image data in a .jpg or .tff file format, then undergoes image
transformation 404 in which a number of image processing and
evaluation functions are performed and the raw image data is
converted into an RGB digital space, an HSV space, and a color
ratio space employing image processing techniques well known in the
art. The RGB-to-HSV transformation, for example, is found at
http://disney.ctr.columbia.edu/jrsthesis/node29.html and
incorporated herein by reference in its entirety. For a raw image
comprising L.times.M pixels, a two-dimensional feature space,
L.times.M matrix of one-by-ten feature vectors comprises the
transformation output. Each one-by-ten feature vector for each
pixel location includes an associated feature value for each pixel
location, (l, m), for example, in the form of 10.times.1 vectors in
the form [R, G, B, H, S, V, GI, RG, RB, GB]. The transformation
output data space and the associated raw image data are saved to a
hard disk as a data file and marked with appropriate identifiers.
The next step is pixel processing 406.
[0063] Referring now to FIG. 5, at the start of pixel processing
500, a user retrieves a data file and views a current image that is
most convenient for manual user observation, for example, the user
views either the raw image, the RGB digital space, HSV space or
color ratio space on the computer monitor. Before proceeding with
the classification, the user selects one of two pixel processing
protocol options at step 510 from the graphical user interface
prompt. One of the protocol options is called "PIXEL LEARN" 512 and
the other is called "PIXEL CLASSIFY" 514. The user chooses "PIXEL
CLASSIFY" 514 if an existing database that has already been trained
is to be employed and applied to the current image for automated
classification. The "PIXEL CLASSIFY" option will be discussed in
detail hereinbelow. The "PIXEL LEARN" 512 option is chosen if the
user does not wish to employ an existing database, an existing
database is not suitable for the tissue sample and image at hand or
a database has not been trained such that it can be applied to the
current image. "PIXEL LEARN" 512 enables the user to create a new
database for the current image. The database that is created under
"PIXEL LEARN" 512 for the current image is stored and may also be
concatenated to a variety of other new or existing, local or global
type databases. Database information management will be discussed
in detail hereinbelow.
[0064] "PIXEL LEARN" is processing along the learning and automatic
classification axis or second dimension processing as opposed to
the variable abstraction axis discussed above. "PIXEL LEARN" and
the second dimension of learning and automatic classification
permits the system to become more intelligent with time as data and
processing results from one or more "PIXEL LEARN" operations are
stored in various databases as will be described hereinbelow. In
general, the learning stage requires interactive processing with
end user participation in which the end user participates by
providing a number, preferably at least three, of ROI and non-ROI
selections for a small number of images. It should be noted that
the number of images is less relevant than the number of ROI and
non-ROI selections in terms of pixels, subimages, and objects. The
reason is that if an image can provide a lot of relevant example
pixels and it is representative of the true operating space, then
only one image may be sufficient. If a particular image has very
little relevant tissues, then more images may be required to create
a learned database. Then, the system takes over and creates
individual trained databases, which are then concatenated to form
local or global databases for use with the "PIXEL CLASSIFY" option
514, which is automated and requires little end user participation.
The "PIXEL LEARN" 512 option will now be discussed in greater
detail.
[0065] In "PIXEL LEARN" 512 under the pixel processing protocol
500, the user selects at least one region of interest (ROI) 516
from the image viewed on the computer monitor using a mouse or some
other pointing device coupled to the computer. A ROI is a desired
target location to be captured via laser capture microdissection
and may include an abnormal cell or cells, malignant or
pre-malignant cell or cells or otherwise a cell or cells of
interest. Preferably, the user selects three or more ROIs. Next,
the user is prompted to select at least one non-ROI 516, or region
that is not of interest for the application. Examples of non-ROIs
include healthy cells or other background matter or clutter. Of
course, the ROIs are not limited to the examples shown herein, such
that a ROI may, of course, be a healthy cell selection.
Nonetheless, the type of ROI selected should be consistent for each
database. Preferably, the user also selects three or more non-ROIs.
Each selection or exemplar made by the user segments the image and
annotates the transformation output vector of the pixel location or
locations of the selection or selections with another value that
identifies the type of selection, ROI or non-ROI, for the pixel
location or locations that were manually selected by the user 518.
The annotation to the vector is a binary value such that, for
example, the number "1" is used for non-ROI selections and the
number "2" is used for ROI selections. Therefore, vectors in a
two-dimensional, L.times.M matrix associated with ROI or non-ROI
selections are augmented to one-by-eleven vectors wherein, for
example, a ROI-selected pixel is denoted by a vector in the form
[R, G, B, H, S, V, GI, RG, RB, GB, 2] and a non-ROI-selected pixel
is denoted by a vector in the form [R, G, B, H, S, V, GI, RG, RB,
GB, 1]. The matrix is stored in the data file associated with the
image 520.
[0066] Alternatively, an N.times.3 array is created wherein N is
the number of ROI and non-ROI manual selections made by the user.
The array includes the pixel location (x, y) and the corresponding
class indices such as 1 for non-ROI and 2 for non-ROI. Hence, a
vector in the array will be of the form [x, y, 2] for a ROI and [x,
y, 1] for a non-ROI segment. In addition to the image file, this
array is another input under the "PIXEL LEARN" processing
option.
[0067] Also, the L.times.M matrix or, alternatively, the N.times.3
matrix together with the image file is stored in a local pixel
database 520. A local database is created and classified in any
number of ways. For example, a local database can consist of data
associated with a particular image or a particular type of tissue
or stain used. Another example of a local pixel database is a
patient-specific database, In general, a variety of local pixel
databases may be created and/or updated.
[0068] The information can also be stored in a global database. A
global database, for example, can be all the information associated
with pixel processing selections regardless of tissue type and
stain. Each time "PIXEL LEARN" is chosen appropriate databases are
automatically or manually concatenated to improve system
performance with time. For example, all data associated with one
tissue sample collected over more than one image can be
concatenated into one database. Another example is the
concatenation of all data associated with a particular tissue type.
In general, a variety of global databases can be created and/or
updated.
[0069] The next step is classification 522 under "PIXEL LEARN."
Referring now to FIG. 6, there is depicted the steps comprising
classification. At the start of classification 600, the first step
is feature extraction 602. In feature extraction, features are
measured and calculated and a matrix of 11.times.1 vectors of
feature values and ROI/non-ROI annotations, such as red (R), green
(G), blue (B), hue (H), saturation (S), intensity or value (V),
gray intensity after RGB-to-gray conversion and histogram
equalization (GI), red/green (RG), red/blue (RB), green/blue (GB),
non-ROI (1), ROI (2) represented by the exemplary vector [R, G, B,
H, S, V, GI, RG, RB, GB, 1], is compiled from each ROI and non-ROI
selection made by the user for a particular data set at the pixel
level.
[0070] The next step is feature ranking 604 in which the pixel
features [R, G, B, H, S, V, GI, RG, RB, GB] are ranked according to
their effectiveness in discriminating ROIs. Feature ranking is
performed using any combinatorial optimization algorithm known in
the art such as add-on or forward-backward ranking algorithm to
first iteratively rank each single feature (singlet), R, G, B, H,
S, V, GI, RG, RB, GB, from best feature to worst feature. Feature
performance is based on the degree of successful detection of a ROI
annotation that is made by the user and reserved for performance
testing. For example, if a user selects four ROIs, two ROI
selections will be reserved to test the performance of features
obtained from the other two ROI selections. Each singlet will be
then ranked accordingly using the combinatorial optimization
algorithm.
[0071] The combinatorial optimization algorithm proceeds to then
iteratively rank pairs of features (doublets). For example, the
first ranked singlet will be paired with another feature and the
performance of both features in combination will be ranked against
other doublets. The combinatorial optimization algorithm proceeds
to then iteratively rank sets of three features (triplets). For
example, the first ranked doublet will be combined with another
feature to form a triplet of features, which will be ranked against
other triplets. The algorithm ranks singlets, doublets, triplets,
and quadruplets and so on until the improvement in performance for
detecting a ROI reaches a predetermined value or point of
diminishing returns at which point a feature singlet, doublet or
triplet is selected. For example, the highest ranked triplet may
provide only a small percent increase in performance over the
highest ranked doublet in which case, the doublet would be the
selected feature combination to be employed in classification under
"PIXEL LEARN." This estimation of the point of diminishing returns
advantageously enhances real-world performance by avoiding
data-overfitting or memorization during learning. That is, the
iterative optimization algorithm automatically finds the minimum
feature dimension that offers the best classification accuracy. The
result of the optimization algorithm is an optimal feature subset
(x), e.g., a singlet, doublet, or triplet optimal feature
selection. The optimal feature subset (x) is stored in a pixel
learn database. The data file may be any of a number of local or
global pixel databases. As additional ROI and non-ROI selections
are made by the user, the optimal feature subset (x) is re-compiled
for the entire data file either local or global. Any local and
global database to which pixel processing results are concatenated
also undergoes the combinatorial optimization algorithm for a
refinement of the resident optimal feature subset.
[0072] The next step is the selection of the classification
algorithm 606 in which an appropriate mapping function is selected
that transforms the optimal feature subset (x) into a discrete
class label (y). The mapping function is implemented in the form of
a classifier. Whereas any classifier known in the art can be
employed, two classification algorithms are implemented for their
simplicity, transparency, and versatility in pattern learning and
discrimination. The two classifiers are multivariate Gaussian (MVG)
and Gaussian mixture model (GMM).
[0073] The GMM models each class-conditional feature probability
density
p(x|y)
[0074] as a mixture of Gaussians as shown below:
p ( x | y = i ) = n = 1 Mg g ni N ( x , .mu. ni , ni ) ##EQU00001##
where ##EQU00001.2## N ( x , .mu. ni , ni ) = 1 ( 2 .pi. ) K ni exp
( - 0.5 ( x - .mu. ni ) t ni - 1 ( x - .mu. ni ) )
##EQU00001.3##
[0075] where M.sub.g represents the number of Gaussian mixtures
that models p(x|y). The default value for the number of Gaussian
mixtures is three; however, the invention is not so limited and any
number of mixtures is within the scope of the invention. The number
of Gaussian mixtures is not greater than the number of exemplars in
each class. For example, if the user makes three ROI selections and
three non-ROI selections, then the maximum possible M.sub.g is
three. Preferably, the number M.sub.g is less than the number of
selections by at least an order of magnitude.
[0076] Further in the above equation, K is the feature dimension,
g.sub.ni is the prior probability associated with the nth mixture
for class i, .mu..sub.ni is the mean vector (centroid) of the nth
mixture for the ith, ( ).sup.t is the transpose operator, and
.SIGMA..sub.ni is the covariance matrix of the nth mixture for the
ith class.
[0077] The MVG classifier is a version of GMM where M.sub.g=1. In
other words, GMM consists of multiple MVGs. MVG parameters consist
of a mean vector and covariance matrix for each class, ROI or
non-ROI. MVG is advantageous because the learning requires no
iterative optimization as in GMM because of the existence of a
closed-form solution. The exponent in the above equation called the
Mahalanobis distance (r) for the ith class is
r.sub.i(x-.mu..sub.i).sup.t.SIGMA..sub.i.sup.-1(x-.mu..sub.i)
[0078] The Mahalanobis distance is computed for the feature vector
of each pixel and the feature vector is classified by measuring the
Mahalanobis distance from the feature vector to the mean vector
(.mu.) and assigning the feature vector to the class for which the
Mahalanobis distance is minimum. The feature vector for each pixel
is annotated with the class label, for example 1 for ROI and 0 for
non-ROI.
[0079] The log-likelihood ratio (LLR) between two classes, ROI and
non-ROI is
M.sub.i-M.sub.j
[0080] The LLR controls the classifier threshold. The LLR threshold
is 0 for neutrality and positive for the ROI class and negative for
the non-ROI class assuming non-ROI-to-ROI ordering.
[0081] MVG models each class-conditional feature distribution in
terms of its mean vector (.mu..sub.ni) and covariance matrix
(.SIGMA..sub.i) that measures the spread around the centroid. For
each unknown feature vector, MVG computes the normalized distance
between the unknown feature vector and the centroid of each
class.
[0082] Selection of the classification algorithm 606 is performed
automatically or manually. The user may select an algorithm
manually from a graphical user interface (GUI). Alternatively,
automatic selection of the classification algorithm is data
dependent and involves a recommendation engine. The recommendation
engine is an algorithm that examines the feature distribution and
recommends the MVG classifier for unimodal feature distributions
and the GMM classifier if the feature distribution is multimodal.
In short, the learning algorithm that is suited for the underlying
feature distribution is selected. FIG. 7 illustrates a simple
example. If the actual feature distribution looks multi-modal as in
FIG. 7a, GMM is the preferred classifier and the number of clusters
determines the number of Gaussian mixtures to be used in
representing each class conditional feature distribution. If the
actual feature distribution looks unimodal as in FIG. 7b, MVG is
the preferred classifier. The recommendation engine also estimates
the most appropriate number of modes for the GMM classifier. As
shown in FIG. 7a, there are two modes for each class.
[0083] As can be seen from above, there are several inputs for the
"PIXEL LEARN" processing option, which will now be summarized. The
inputs include the image file, which is typically a .jpg or .tif
file type. Another input is the N.times.3 array for identifying
pixel locations and their corresponding classification indices
selected manually by the user for N number of selections. Another
input includes various processing parameters. These processing
parameters are stored in a 1.times.5 row vector wherein the first
parameter is the minimum region area in pixels. The user will input
the minimum region area based on a number of considerations
including laser spot constraints and minimum cell size for single
cell capture capability. The second parameter input is the
classifier selection. For example, if the user manually inputs or
if the recommendation engine automatically selects MVG as the
classifier of choice, the second parameter will include a
classifier identification such as "1" for MVG and "2" for GMM. The
third parameter is a classifier parameter. This classifier
parameter is dependent on the type of classifier selected. For
example, for the MVG classifier, the LLR offset is the third
parameter input where a positive LLR favors ROI detection. For the
GMM classifier, the third parameter input is the number of Gaussian
mixtures. The fourth parameter of the 1.times.5 row vector of
processing parameters is the region of influence (M). The region of
influence is a block of M.times.M pixels centered on the pixel
location for each pixel location in the N.times.3 pixel-training
array. The block is used to derive the pixel-learn database and is
manually inputted or pre-selected for maximum robustness. The last
element in the processing parameters is a flag that is reserved for
debugging statements. For example, if this parameter is set to "1",
then debug statements will be written to a text file. Of course,
another input is a file name that is used to concatenate pixel
learn data for tissue types for example. The file name input
includes both path and file name.
[0084] The output 608 of "PIXEL LEARN" processing is a binary image
map for a two-class problem ("1" for non-ROI and "2" for ROI) that
is displayed on the computer monitor for user inspection. The
output also provides region labeling in which connected pixels of
the same class arc aggregated into a region and labeled as a ROI or
non-ROI region. Each region is numbered and the region label and
number are stored in one or more databases. The labeling can be
performed in a variety of ways, and color is used to denote a
region identification number for easy user identification of
regions. FIG. 8a shows the results of pixel processing and the
labeling of regions 802 of the binary image map are shown in FIG.
8b.
[0085] Additionally, a post-processing step 610 is included to
identify potential regions of interest for further processing. For
example, in one variation, the binary image map undergoes a
post-processing step that determines the size of each region 802.
The determination of size of each region assists in the preliminary
rejection of a region based on a user-specified size threshold. For
example, if the region size is smaller than an average cell size,
that region may be discarded from further processing as being
considered undesirable for the application such as LCM.
[0086] The output of the "PIXEL LEARN" processing option is stored
in a local global database 612. The output of "PIXEL LEARN"
includes the original image, the labeled image, a boundary
structure summarizing classification results, and any error message
that may be applicable. The boundary structure that summarizes
classification results consists of several fields including a list
of pixel locations (x, y), an interior flag, the region area in
pixels, the length of the perimeter of a region and the number of
boundary pixels. The interior flag is set to one (1) if the
boundary of a region belongs to the interior of an object and set
to zero (0) otherwise. The list of pixel locations may be sorted,
for example by region and by classification. In order to save
memory, a polygonal approximation of the external boundary is
performed as well.
[0087] Database creation and management will now be discussed. As
shown in FIG. 4, pixel processing 406 is in electrical
communication with various local and global databases 414. This
communication with various databases under the "PIXEL LEARN"
protocol is illustrated in FIG. 9. Communication with various
databases under the "PIXEL CLASSIFY" protocol will be discussed
later herewith.
[0088] As discussed above, if the user selects pixel processing
900, the user will be prompted at 902 by the GUI to select either
the "PIXEL LEARN" 906 or "PIXEL CLASSIFY" 907. If the user does not
select pixel processing 900, processing will be directed at 904 to
proceed under subimage, object and/or LCM processing as discussed
above with respect to FIG. 4. Under "PIXEL LEARN" 906, a local
database for a particular image is created 908. For example, local
database A is created from pixel processing of image A, which was
obtained from a first pancreatic tissue sample for example. The
next step is a query at 910 to create a new global database. If
yes, a new global database is created 912. Continuing with the
example, this new global database, DB1, comprises all the
information associated with local database A. This new global
database, for example, can be named "pixel data for the first
pancreatic tissue sample" in which all the data from all the images
of the first tissue sample will be stored. A query to update an
existing global database 914 is next. If yes, an existing global
database is updated 916 and once again the query 914 is posed to
update another existing global database. Continuing with our
example, there are no existing global databases to be updated. If
no, the query 804 is posed to create another global database. In
our example, another global database, DB2 is created 912, for
example, and named "pixel data for all pancreatic tissue samples"
in which all pixel data associated with all pancreatic tissue
samples will be stored. Since in the example there are no existing
databases to update 914, 918 and no new global databases to create
910, a query for pixel processing 900 is posed. If pixel processing
900 is desired for another image, the user selects between the
"PIXEL LEARN" and "PIXEL CLASSIFY" at 902 once again for the next
image to be processed. If pixel processing 900 is not desired, the
classification may stop after pixel processing or proceed with
subimage, object processing and/or LCM 904. To illustrate the
concatenation of databases in this example, another image B but of
different location of the same first tissue sample is obtained and
processed under "PIXEL LEARN" 906. A local database, database B, is
created for image B at 908. At 910, the user is prompted to create
a new global database. Using the example, two global databases, DB1
and DB2, have already been created. If it is not desired to create
new global databases, the next step is a query 918 to update
existing global databases. An existing global database is updated
at 920. In the example, data from image B that is resident in
database B is used to update existing database DB1. It makes sense
to update DB1 since image B is of the first pancreatic tissue
sample and data from image B is concatenated to global database
DB1, which is called "pixel data for the first pancreatic tissue
sample." Now global database DB1 comprises data of local databases
A and B. Once again at 918, another existing database may be
updated. Using the example, DB2, which is the database for all
pancreatic tissue samples, can be selected and updated 920.
Likewise DB2 is concatenated with data from database B such that
DB2 comprises data of both A and B. Again, a query to proceed with
pixel processing is made at 900. As can be seen in the example,
with each "PIXEL LEARN" process an updated database grows as more
and more parameters are acquired. Updating a database 612 includes
not only a concatenation of data but also a re-evaluation of the
newly aggregated data. This re-evaluation is performed
automatically by an update module and includes a re-ranking of
features and a re-selection of the optimal feature subset. Other
actions performed in the updating of a database include dimension
reduction and an estimation of learning algorithm parameters. The
updating of a database is repeated with each new concatenation so
that the underlying image algorithms become more intelligent as
they age. The databases are updated to reflect the latest changes
in tissue samples or staining methods such that when they are
retrieved they are suitable for use with processing at the same or
next level of abstraction. From "PIXEL LEARN" the user may select
to proceed with LCM if it so desired.
[0089] The "PIXEL CLASSIFY" option will now be discussed. "PIXEL
CLASSIFY" employs data accumulated and stored in various databases
from one or more learning stages such as the "PIXEL LEARN" stage.
As discussed above, during "PIXEL LEARN" the user provides truth
annotations, for example ROI and non-ROI selections for
representative images using GUI controls. Then, the system
discriminates ROIs from non-ROIs using classifiers. During the
learning stage, all the parameters and feature values associated
with pattern learning are stored for each image being investigated
in one or more database such as local and global databases. The
data from "PIXEL LEARN" is also concatenated to one or more global
databases as discussed above with respect to FIG. 9. During "PIXEL
CLASSIFY" the system utilizes the learned parameters stored in one
or more databases to perform automatic classification of regions as
being ROIs or non-ROIs. As the databases grow with time as more
data is acquired during "PIXEL LEARN," "PIXEL CLASSIFY" improves
with age. The "PIXEL CLASSIFY" operating mode is designed for
high-throughput batch processing with the human operator inspecting
the processed results for final decision.
[0090] Referring back to FIG. 5, after "PIXEL CLASSIFY" is selected
as the pixel processing option at 510, "PIXEL CLASSIFY" will start
514 by prompting the user via the GUI to load an appropriate
database 524 if one is available. If no appropriate database is
available, the user will be directed to start "PIXEL LEARN" 512. A
pull-down menu on the GUI provides a list of appropriate local and
global databases to be selected 526 by the user under "PIXEL
CLASSIFY." Typically, the available databases are local or global
databases created during "PIXEL LEARN" and matched for the
particular image at hand. For example, if the image is a pancreatic
tissue sample, the user may select any local or global pancreatic
tissue database at the pixel level of abstraction at 526. The next
step is to run the classification 528 under "PIXEL CLASSIFY."
[0091] Running the classification under "PIXEL CLASSIFY" is
depicted in FIG. 10. After the start of "run classification" 1000,
the database that was selected at 526 is accessed 1002 for critical
algorithmic parameters and for the optimal feature subset. The
selected database has already been updated at 520. Hence, the
database has undergone a ranking or re-ranking of features and
selection or re-selection of the optimal feature subset. Therefore,
these values are ready for use with the classification algorithm
that is selected at 1004. In "PIXEL CLASSIFY," the classification
algorithm is selected automatically. Automatic selection of the
classification algorithm is data dependent and involves a
recommendation engine. The recommendation engine is an algorithm
that examines the feature distribution and recommends the MVG
classifier for unimodal feature distributions and the GMM
classifier if the feature distribution is multimodal. The
recommendation engine also estimates the most appropriate number of
modes for the GMM classifier.
[0092] The output 1006 of "PIXEL CLASSIFY" is the same as the
output of "PIXEL LEARN." The output 1006 of "PIXEL CLASSIFY"
processing is a binary image map for the two-class problem ("1" for
non-ROI and "2" for ROI) that is displayed on the computer monitor
for user inspection. The output also provides region labeling in
which connected pixels of the same class are aggregated into a
region and labeled as a ROI or non-ROI region. Each region is
numbered and the region label and number are stored in one or more
databases. The labeling can be performed in a variety of ways, and
color is used to denote a region identification number for easy
user identification of regions.
[0093] The final step of "run classification" is post-processing
1008. Post-processing 1008 under "PIXEL CLASSIFY" is the same as
post-processing 610 under "PIXEL LEARN." The post-processing step
1008 identifies potential regions of interest for further
processing. For example, in one variation, the binary image map
undergoes a post-processing step that determines the size of each
region. The determination of size of each region assists in the
preliminary rejection of a region based on a user-specified size
threshold. For example, if the region size is smaller than an
average cell size, that region may be discarded from further
processing as being considered undesirable for the application such
as LCM.
[0094] The output of "PIXEL CLASSIFY" includes the original image,
the labeled image, a boundary structure summarizing classification
results, and any error message that may be applicable. The boundary
structure that summarizes classification results consists of
several fields including a list of pixel locations (x, y), an
interior flag, the region area in pixels, the length of the
perimeter of a region and the number of boundary pixels. The
interior flag is set to one (1) if the boundary of a region belongs
to the interior of an object and set to zero (0) otherwise. The
list of pixel locations may be sorted, for example by region and by
classification. In order to save memory, a polygonal approximation
of the external boundary is performed as well.
[0095] Upon viewing the output of either the "PIXEL LEARN" or
"PIXEL CLASSIFY" options, the user determines whether to proceed to
LCM along line 416 in FIG. 4 or along line 418 to the next level of
abstraction, subimage processing or object processing. If the user
determines that the output binary image map is satisfactory and
does not include any false accepts, then the user selects LCM from
the GUI. If the tissue slide is loaded in the LCM device, the
processor will automatically position the laser at coordinates
corresponding to ROIs from the binary image map. Once positioned,
the laser is activated for LCM and the ROIs are transferred to the
LCM transfer film and removed from the tissue sample. If the user
determines that the output binary image map is not satisfactory due
to a more complex image, for example, the user then selects from
the GUI to proceed to subimage processing 408. Alternatively, the
user may select to bypass subimage and proceed to object processing
410 prior to LCM 412.
[0096] Referring now to FIG. 11, subimage processing will now be
discussed. Subimage processing overcomes difficulties associated
with the output of pixel processing. For example, subimage
processing will differentiate between two classes, ROI and non-ROI,
with similar color by extracting additional features related to
texture over a swath of the image that is larger than a single
pixel. At the start of subimage processing 1100, if a data image
file is not already loaded from the previous step of pixel
processing, the user is prompted by the GUI to retrieve a image
data file corresponding to the current image that has undergone
pixel processing for user observation 1101. The image, typically in
.jpg or .tff file format is inputted along with the labeled image
from pixel processing at 1101. At the same time, in one variation,
the pixel processing data file corresponding to the image at hand
is saved as a subimage processing data file for the same image.
Subimage processing stores data to this newly created subimage
processing data file. Typically, the user views the raw image of
the tissue sample on the computer monitor. Before proceeding with
the classification, the user selects one of two subimage processing
protocol options at 1102 from the GUI. One of the subimage
processing protocols is "SUBIMAGE LEARN" 1104 and the other is
called "SUBIMAGE CLASSIFY" 1106. The user chooses "SUBIMAGE
CLASSIFY" 1106 if an existing database is to be employed and
applied to the current image for automated classification. The
"SUBIMAGE CLASSIFY" option 1106 will be discussed in detail
hereinbelow. The "SUBIMAGE LEARN" 1104 option is chosen if the user
does not wish to employ an existing database and apply it to the
current image. "SUBIMAGE LEARN" 1104 enables the user to create a
new subimage database for the current image. The subimage database
for the current image is concatenated to a variety of other new or
existing, local or global type databases.
[0097] "SUBIMAGE LEARN" 1104 is processing along the learning and
automatic classification axis or second dimension processing as
opposed to the variable abstraction axis discussed above. "SUBIMAGE
LEARN" and the second dimension of learning and automatic
classification permits the system to become more intelligent with
time as data and processing results from the "SUBIMAGE LEARN" stage
are stored in various databases.
[0098] In general, "SUBIMAGE LEARN," as with any of the learning
stage, requires interactive processing with end user participation
in which the end user participates by providing a predetermined
number, preferably at least three, of ROI and non-ROI selections
all at the subimage level of abstraction. Then, the system takes
over and creates individual trained databases, which are then
concatenated to form local or global databases for use with the
"SUBIMAGE CLASSIFY" option 1106, which is automated and requires
little end user participation. The "SUBIMAGE LEARN" 1104 option
will now be discussed in greater detail.
[0099] In "SUBIMAGE LEARN" 1104 under the subimage processing
protocol 1100, the user selects at least one subimage region of
interest (ROI) 1108 from the image viewed on the computer monitor
using a mouse or some other pointing device coupled to the
computer. Preferably, the user selects three or more subimage ROIs.
The user is also prompted to select at least one subimage non-ROI
1108. Preferably, the user also selects three or more subimage
non-ROIs. Each subimage selection is a selection of a region larger
than one pixel. Generally, the subimage region is equal to or
larger than a predetermined minimum size that is inputted by the
user. This predetermined minimum size for a subimage selection is
preferably 4 to approximately 32 pixels. The predetermined minimum
can be tailored according to various tissue types undergoing study.
In one variation, the user makes a subimage selection using a GUI
interface to draw a polygonal shape. The polygonal shape is drawn
by clicking on an area of the image with a pointing device and then
dragging the pointer to designate a polygonal subimage ROI or
non-ROI segment. The maximum size of the ROI and non-ROI subimage
segment is at the user's discretion so long as it is a subset of
the entire image and does not overlap with other ROI or non-ROI
polygonal region selections. Once a subimage ROI is designated on
the image, for example, by a polygonal shape, that designation is
indicated in a color that is different from a subimage non-ROI
designation on the same image for easy user identification. The
user-specified subimage regions are non-overlapping regions as
shown in FIG. 12. For example, the subimage ROI polygons 1202 are
magenta in color and the subimage non-ROI polygons 1204 are
yellow.
[0100] With the subimage selections, the image data file is
annotated with the additional information obtained from the
subimage selections at 1110. Each subimage selection or exemplar
annotates the pixel output vector of all the pixel locations
associated with the subimage region with another value that
identifies the type of selection, ROI or non-ROI, for the pixel
locations included in the subimage region. This annotation is
performed for the entire L.times.M matrix for the image. The
annotation to each vector is a binary value such that, for example,
the number "1" is used for non-ROI selections and the number "2" is
used for non-ROI selections. The matrix is stored in the subimage
data file associated with the image 1112.
[0101] The next step is to run the classification 1114. Referring
now to FIG. 13, there is depicted the steps comprising run
classification under "SUBIMAGE LEARN." At the start of
classification 1300, the first step is feature extraction 1302 at
the subimage level. Feature extraction 1302 at the subimage level
differs from feature extraction at the pixel level performed during
pixel processing. In subimage processing, feature extraction uses
pixel results to reduce the problem dimension by extracting a
different set of features such as texture-related information at a
subimage level selected by user. The following table lists the
types of features that are extracted at the subimage level.
TABLE-US-00001 TABLE OF SUBIMAGE-LEVEL FEATURES 1 Mean intensity of
each subimage 2 Standard deviation of subimage pixels 3 Coefficient
of variation 4 (Maximum pixel intensity - minimum pixel
intensity)/standard deviation 5 The fraction of low-frequency
energy as defined by the discrete cosine transformation (DCT) 6 DC
energy fraction, which is the mean energy of each subimage (mean of
R, G, and B) 7 Mean/standard deviation of DCT coefficients in the
LL sub-band quadrant (L = low frequency, H = high frequency) 8
Mean/standard deviation of DCT coefficients in the LH sub-band
quadrant 9 Mean/standard deviation of DCT coefficients in the HL
sub-band quadrant 10 Mean/standard deviation of DCT coefficients in
the HH sub-band quadrant 11 Mean of vertical difference in
intensity 12 Mean of horizontal difference in intensity 13 Standard
deviation of vertical difference in intensity 14 Standard deviation
of horizontal difference in intensity 15 Coefficient of variation
of vertical difference in intensity 16 Coefficient of variation of
horizontal difference in intensity
[0102] The above features are selected to characterize each
subimage in terms of trend, texture, normalized variation and
normalized color.
[0103] The above features are repeated as applied to the RGB space
such that the final analysis results are based on 48 features.
These 48 features include the above 1-16 features derived or
calculated from the red space, the above features repeated and
derived or calculated from the green space (17-32), and the above
features repeated and derived or calculated from the blue space
(33-48). In one variation, in addition to the above features being
repeated for the RGB space, they are also repeated for the HSV
space, such that the above features are repeated for the hue space,
saturation space and intensity space. As a result, a total of 96
features are extracted in such variation.
[0104] In yet another variation, the following table lists another
set of features that are extracted at the subimage level.
TABLE-US-00002 TABLE OF SUBIMAGE-LEVEL FEATURES 1 Mean 2 Standard
deviation 3 Mean color differences in the horizontal direction 4
Standard deviation of color differences in the horizontal direction
5 Mean color differences in the vertical direction 6 Standard
deviation of color differences in the vertical direction
[0105] The above features are repeated as applied to the RGB space
such that the final analysis results are based on 18 features. The
above 18 features maximize computational speed while at the same
time prove to be reasonably effective as compared to the previous
48 features. In one variation, the user selects which set of
features from a library set to employ for a given situation, which
of course is dependent on the complexity of the image to be
classified.
[0106] The features are extracted from the image, compiled in
vector form and stored with the associated subimage data file. The
next step is feature ranking 1304 in which the subimage features
are ranked according to their effectiveness in discriminating ROIs.
Feature ranking is performed in substantially the same way as in
pixel processing by using any combinatorial optimization algorithm
known in the art such as add-on or forward-backward ranking
algorithm to first iteratively rank each single subimage feature
(singlet) from best feature to worst feature. Feature performance
is based on the degree of successful detection of subimages that
fall into a ROI-annotated region that is made by the user and
reserved for performance testing. For example, if a user selects
four ROI subimage regions, two ROI selections will be reserved to
test the performance of features obtained from the other two ROI
selections. Each singlet will be then ranked accordingly using the
combinatorial optimization algorithm.
[0107] The combinatorial optimization algorithm proceeds to then
iteratively rank pairs of features (doublets). For example, the
first ranked singlet will be paired with another feature and the
performance of both features in combination will be ranked against
other doublets. The combinatorial optimization algorithm proceeds
to then iteratively rank sets of three features (triplets). For
example, the first ranked doublet will be combined with another
feature to form a triplet of features, which will be ranked against
other triplets. The algorithm ranks singlets, doublets, triplets,
quadruplets and so on until the improvement in performance for
detecting a ROI reaches a predetermined value or point of
diminishing returns at which point a feature singlet, doublet or
triplet is extracted. For example, the highest ranked triplet may
provide only a small percent increase in performance over the
highest ranked doublet in which case, the doublet would be the
selected feature combination to be employed in the next step in
classification. This estimation of the point of diminishing returns
advantageously enhances real-world performance by avoiding
data-overfitting or memorization during learning. The result of the
optimization algorithm is a subimage feature subset (x) e.g., a
singlet, doublet, or triplet optimal subimage feature selection.
The optimal subimage feature subset (x) is stored in the subimage
data file. The data file may be any of a number of local or global
databases.
[0108] The next step is the selection of the classification
algorithm 1306 in which an appropriate mapping function is selected
that transforms the optimal subimage feature subset (x) into a
discrete class label (y). The mapping function is implemented in
the form of a classifier. Classification algorithm selection 1306
is performed in the same way as in pixel processing. Selection of
the classification algorithm 1306 is performed automatically or
manually. The user may select a classification algorithm manually
from a GUI interface that provides a list of any number of suitable
algorithms. Typically, the selection involves a choice between GMM
and MVG. Alternatively, automatic selection of the classification
algorithm is data dependent and involves a recommendation engine.
The recommendation engine is an algorithm that examines the
subimage feature distribution and recommends the MVG classifier for
unimodal feature distributions and the GMM classifier if the
feature distribution is multimodal. If the actual feature
distribution looks multi-modal, GMM is the preferred classifier and
the number of clusters determines the number of Gaussian mixtures
to be used in representing each class conditional feature
distribution. If the actual feature distribution looks unimodal,
MVG is the preferred classifier. The recommendation engine also
estimates the most appropriate number of modes for the GMM
classifier.
[0109] As can be seen from above, there are several inputs for the
"SUBIMAGE LEARN" processing option, which will now be summarized.
The inputs include the image file, which is typically a .jpg or
.tif file type, and the labeled image from pixel processing.
Another input is the N.times.3 array for identifying pixel
locations and their corresponding classification indices selected
manually by the user for all of the pixels in each subimage
selection. Another input includes various processing parameters.
These processing parameters are stored in a 1.times.5 row vector
wherein the first parameter is the minimum subimage region size in
pixels. The user will provide the minimum region area based on a
number of considerations including laser spot constraints and
minimum cell size for single cell capture capability. The second
parameter is the actual size of the subimage selected by the user.
The third parameter input is the classifier selection. For example,
if the user manually inputs or if the recommendation engine
automatically selects MVG as the classifier of choice, the second
parameter will include a classifier identification such as "1" for
MVG and "2" for GMM. The fourth parameter is a classifier
parameter. This classifier parameter is dependent on the type of
classifier selected. For example, for the MVG classifier, the LLR
offset is the fourth parameter input where a positive LLR favors
ROI detection. For the GMM classifier, the fourth parameter input
is the number of Gaussian mixtures. The fifth and last element in
the processing parameters is a flag that is reserved for debugging
statements. For example, if this parameter is set to "1", then
debug statements will be written to a text file. Of course, another
input is a file name, which is used to concatenate subimage learn
data for particular tissue types. The file name input includes both
path and file name.
[0110] The output 1308 of "SUBIMAGE LEARN" processing is a binary
subimage map for a two-class problem ("1" for non-ROI and "2" for
ROI) that is displayed on the computer monitor for user inspection.
The output also provides region labeling in which connected pixels
of the same class are aggregated into a region and labeled as a ROI
or non-ROI region. Each region is numbered and the region label and
number are stored in one or more databases. The labeling can be
performed in a variety of ways, and color is used to denote a
region identification number for easy user identification of
regions. FIG. 14a shows the results of subimage processing. FIG.
14b shows the results of pixel processing alone for the same
image.
[0111] Additionally, a post-processing step 1310 is included to
identify potential regions of interest for further processing. For
example, in one variation, the binary image map undergoes a
post-processing step that determines the size of each region. The
determination of size of each region assists in the preliminary
rejection of a region based on a user-specified size threshold. For
example, if the region size is smaller than an average cell size,
that region may be discarded from further processing as being
considered undesirable for the application such as LCM. Also,
during the post-processing step 1310, a false accept rejection is
performed. The results of subimage processing and pixel processing
are combined to create a cleansed binary image map.
[0112] The output of the "SUBIMAGE LEARN" processing option is
stored in a local database. The output of "SUBIMAGE LEARN" includes
the original image, a further refined and labeled image, a boundary
structure summarizing classification results, and any error message
that may be applicable. The boundary structure that summarizes
classification results consists of several fields including a list
of pixel locations (x, y) and their corresponding classification,
an interior flag, the region area in pixels, the length of the
perimeter of a region and the number of boundary pixels. The
interior flag is set to one (1) if the boundary of a region belongs
to the interior of an object and set to zero (0) otherwise. The
list of pixel locations may be sorted, for example by region and by
classification.
[0113] Database creation and management 1312 is the same as that
discussed above with respect to pixel processing and described in
FIG. 9. Basically, database management and creation 1312 for
subimage processing is the same as shown in FIG. 9 wherein the word
"pixel" is replaced with the word "subimage." As shown in FIG. 4,
subimage processing 408 is in electrical communication with various
local and global databases 414. As additional ROI and non-ROI
selections are made by the user when in the "SUBIMAGE LEARN"
processing protocol, the optimal subimage feature subset (x) is
re-compiled for the entire data file either local or global. Any
local and global database to which subimage processing results are
concatenated also undergoes the combinatorial optimization
algorithm for a refinement of the resident optimal subimage feature
subset.
[0114] The "SUBIMAGE CLASSIFY" option will now be discussed.
"SUBIMAGE CLASSIFY" employs data accumulated and stored in various
databases from one or more learning stages such as the "SUBIMAGE
LEARN" stage. As discussed above, during "SUBIMAGE LEARN" the user
provides truth annotations, for example ROI and non-ROI subimage
selections for representative images using GUI controls. Then, the
system discriminates ROIs from non-ROIs using classifiers. During
the learning stage, all the parameters and feature values
associated with pattern learning are stored for each image being
investigated in one or more database such as local and global
databases. The data from "SUBIMAGE LEARN" is also concatenated to
one or more global databases as discussed above. During "SUBIMAGE
CLASSIFY" the system utilizes the learned parameters stored in one
or more subimage databases to perform automatic classification of
regions as being ROIs or non-ROIs. As the databases grow with time
as more data is acquired during "SUBIMAGE LEARN," "SUBIMAGE
CLASSIFY" improves with age. The "SUBIMAGE CLASSIFY" operating mode
is designed for high-throughput batch processing with the human
operator inspecting the processed results for final decision.
[0115] Referring back to FIG. 11, after "SUBIMAGE CLASSIFY" is
selected as the subimage processing option at 1102, "SUBIMAGE
CLASSIFY" will start 1106 by prompting the user via the GUI to load
an appropriate database 1116 if one is available. If no appropriate
database is available, the user will be directed to start "SUBIMAGE
LEARN" 1104. A pull-down menu on the GUI provides a list of
appropriate local and global databases to be selected by the user
under "SUBIMAGE CLASSIFY" 1118. Typically, the available databases
are ones created during "SUBIMAGE LEARN" and matched for the
particular image at hand. For example, if the image is a pancreatic
tissue sample, the user may select any subimage local or global
pancreatic tissue database. Of course, the image of the tissue
sample is also loaded and displayed on the monitor when the
database is selected. The next step is to run the classification
under "SUBIMAGE CLASSIFY" 1120.
[0116] Running the classification under "SUBIMAGE CLASSIFY" is
depicted in FIG. 15. After the start of "run classification" 1500,
the database that was selected at 1118 is accessed 1502 for
critical algorithmic parameters and for the optimal subimage
feature subset. The selected subimage database has already been
updated at 1112. Hence, the database has undergone a ranking or
re-ranking of features and selection or re-selection of the optimal
subimage feature subset. Therefore, these values are ready for use
with the classification algorithm that is selected at 1504.
[0117] In "SUBIMAGE CLASSIFY," the classification algorithm is
selected automatically. Automatic selection of the classification
algorithm is data dependent and involves a recommendation engine.
The recommendation engine is an algorithm that examines the feature
distribution and recommends the MVG classifier for unimodal
subimage feature distributions and the GMM classifier if the
feature distribution is multimodal. The recommendation engine also
estimates the most appropriate number of modes for the GMM
classifier.
[0118] The output 1506 of "SUBIMAGE CLASSIFY" is the same as the
output of "SUBIMAGE LEARN." The output of "SUBIMAGE CLASSIFY"
processing is a binary image map for the two-class problem ("1" for
non-ROI and "2" for ROI) that is displayed on the computer monitor
for user inspection. The output also provides region labeling in
which connected pixels of the same class are aggregated into a
region and labeled as a subimage ROI or non-ROI region. Each region
is numbered and the region label and number are stored in one or
more databases. The labeling can be performed in a variety of ways,
and color is used to denote a region identification number for easy
user identification of regions.
[0119] The final step of "run classification" is post-processing
1508. Post-processing under "SUBIMAGE CLASSIFY" is substantially
the same as post-processing under "SUBIMAGE LEARN." The
post-processing step identifies potential regions of interest for
further processing. For example, in one variation, the binary image
map undergoes a post-processing step that determines the size of
each region. The determination of size of each region assists in
the preliminary rejection of a region based on a user-specified
size threshold. For example, if the region size is smaller than an
average cell size, that region may be discarded from further
processing as being considered undesirable for the application such
as LCM.
[0120] The output of "SUBIMAGE CLASSIFY" includes the original
image, the labeled image, a boundary structure summarizing
classification results, and any error message that may be
applicable. The boundary structure that summarizes classification
results consists of several fields including a list of pixel
locations (x, y), an interior flag, the region area in pixels, the
length of the perimeter of a region and the number of boundary
pixels. The interior flag is set to one (1) if the boundary of a
region belongs to the interior of an object and set to zero (0)
otherwise. The list of pixel locations may be sorted, for example
by region and by classification. In order to save memory, a
polygonal approximation of the external boundary is performed as
well.
[0121] Upon viewing the output of either the "SUBIMAGE LEARN" or
"SUBIMAGE CLASSIFY" options, the user determines whether to proceed
to LCM along line 416 in FIG. 4 or along line 418 to the next level
of abstraction which is object processing. If the user determines
that the output binary image map is satisfactory and does not
include any false accepts, then the user selects LCM from the GUI.
If the tissue slide is loaded in the LCM device, the processor will
automatically position the laser at coordinates corresponding to
ROIs from the binary image map. Once positioned, the laser is
activated for LCM and the ROIs are automatically transferred to the
LCM transfer film and removed from the tissue sample. If the user
determines that the output binary image map is not satisfactory due
to a more complex image, for example, the user then selects from
the GUI to proceed to object processing 410.
[0122] Referring now to FIG. 16, object processing which is also
called region processing will now be discussed. Object processing
overcomes difficulties associated with the output of pixel
processing or subimage processing by utilizing texture, morphology
and gradient information at an object level. At the start of object
processing 1600, if a data image file is not already loaded from
the previous step of pixel processing or subimage processing, the
user is prompted by the GUI to retrieve an image data file
corresponding to the current image that has undergone pixel
processing or subimage processing 1602. The image, typically in
.jpg or .tff file format, is inputted along with the labeled image
from pixel processing or subimage processing at 1602. At the same
time, in one variation, the pixel processing or subimage processing
data file corresponding to the image at hand is saved as an object
processing data file for the same image. Object processing stores
data to this newly created object processing data file. Typically,
the user views the raw image or the labeled image of the tissue
sample on the computer monitor. The user selects one of two object
processing protocol options at 1604 from the GUI. One of the object
processing protocols is "OBJECT LEARN" 1606 and the other is called
"OBJECT CLASSIFY" 1608. The user chooses "OBJECT CLASSIFY" 1608 if
an existing object database is to be employed and applied to the
current image for automated classification. The "OBJECT CLASSIFY"
option 1608 will be discussed in detail hereinbelow. The "OBJECT
LEARN" 1606 option is chosen to create a new object database or if
the user does not wish to employ an existing database and apply it
to the current image. "OBJECT LEARN" 1606 enables the user to
create a new object database for the current image. The object
database for the current image is concatenated to a variety of
other new or existing, local or global type databases.
[0123] "OBJECT LEARN" 1606 is processing along the learning axis or
second dimension processing as opposed to the variable abstraction
axis discussed above. "OBJECT LEARN" and the second dimension of
learning and automatic classification permit the system to become
more intelligent with time as data and processing results from the
"OBJECT LEARN" stage are stored in various databases. In general,
"OBJECT LEARN," as with any of the learning stage, requires
interactive processing with end user participation in which the end
user participates by providing input at the object level of
abstraction. Then, the system takes over and creates individual
object-level databases, which are then concatenated to form local
or global object-level databases for use with the "OBJECT CLASSIFY"
option 1608, which is automated and requires little end user
participation. The "OBJECT LEARN" 1606 option will now be discussed
in greater detail.
[0124] In "OBJECT LEARN" 1606 under the object processing protocol
1600, the user inspects the labeled image from the pixel or
subimage processing protocol 1610. Upon inspection of the image the
user will make one or more selections 1612. If the user is
satisfied with the results from pixel or subimage processing, then
no ROI selections will be made and the object processing protocol
will automatically assign all of the ROIs appearing on the labeled
image for the application at hand such as segmentation 1614. For
example, all the labeled ROIs of the labeled image space will be
assigned for laser capture microdissection by the processor and the
LCM device will be directed by the processor to execute system
functions to extract the ROIs from the tissue sample. If the user
selects at least one ROI using a mouse or some other pointing
device coupled to the computer, the GUI control will prompt the
user at 1616 whether a selection of at least one non-ROI is also
desired. A selection of a non-ROI includes selecting a ROI-labeled
region that is a false-accept using a mouse or some other pointing
device coupled to the computer. A selection of a non-ROI region is
a region that is mistakenly labeled as a ROI by the previous
subimage or pixel processing protocol but in reality is discerned
by the user to be clutter, background or basically a non-ROI. If
the user does not select any non-ROIs, then the object processing
protocol will automatically assign all of the ROIs appearing on the
labeled image at 1618 for the application at hand such as
segmentation. For example, all the ROIs will be assigned for laser
capture microdissection by the processor and the LCM device will be
directed by the processor to execute system functions to extract
the ROIs from the tissue sample. If the user selects at least one
non-ROI at 1616, then the object processing protocol under "OBJECT
LEARN" will commence run classification 1620. The non-ROI selection
by the user will constitute an input and stored in an object data
file as a non-ROI array that associates a particular labeled region
as a false-alarm. Also, a ROI selection by the user will constitute
an input and stored in the same object-level data file as a ROI
array that associates a particular labeled region as an object ROI.
Once an object ROI is designated on the image, that designation is
indicated in a color that is different from an object level non-ROI
designation on the same image for easy user identification.
[0125] With the object-level selections, the object data file is
annotated with the additional information obtained from the object
level selections. Each object-level selection creates an
object-level ROI membership array and an object-level, non-ROI
membership array also called a false-alarm membership array, both
of which are stored with the associated object data file as
input.
[0126] The next step is to run the classification 1620. Referring
now to FIG. 17, there is depicted the steps comprising run
classification 1620 in object processing under the "OBJECT LEARN"
option 1606. At the start of run classification 1700, the first
step is feature extraction 1702 at the object level. Feature
extraction 1702 at the object level extracts features that are
different from the features extracted at the pixel or subimage
level. The following table lists the types of features that are
extracted at the object level.
TABLE-US-00003 TABLE OF OBJECT-LEVEL FEATURES 1 Major axis length 2
Minor axis length 3 Orientation: the angle between the x-axis and
the major axis length 4 Solidity: the fraction of pixels in the
object convex hull that are also in the object 5 Eccentricity (e):
The ratio of the distance between the foci to the major axis length
(e = 0 if a circle, e = 1 if a line) 6 Filled area: size of an
object 7 Euler number: Euler number of a binary image for each ROI
based on William Pratt's book titled Pratt, William K., Digital
Image Processing, New York: John Wiley & Sons, Inc., 1991. p.
633, and incorporated herein by reference in its entirety. 8
Extent: similar to solidity, except using a bounded box, not a
convex hull 9 Equivalent diameter: diameter (2r) of a circle, where
filled area = 4.pi.r.sup.2 10 Global texture (3): standard
deviation alone 11 Difference texture (6): similar in concept to
subimage processing (standard deviation alone) 12 Gradient features
(2): compute a distance set between centroid and boundary pixels at
45-degree increments
[0127] The features are extracted from the image, compiled in
vector form and stored with the associated object data file. The
next step is feature ranking and selection 1704 in which the
object-level features are ranked according to their effectiveness
in discriminating ROIs. Feature ranking is performed in the same
way as in pixel and subimage processing by using any combinatorial
optimization algorithm known in the art such to first iteratively
rank each single object-level feature (singlet) from best feature
to worst feature. Feature performance is based on the degree of
successful detection of a ROI-annotated region that is made by the
user and reserved for performance testing. For example, if a user
selects four ROI object-level regions, two ROI selections will be
reserved to test the performance of features obtained from the
other two ROI selections. Each singlet will be then ranked
accordingly using the combinatorial optimization algorithm.
[0128] The combinatorial optimization algorithm proceeds to then
iteratively rank pairs of features (doublets). For example, the
first ranked singlet will be paired with another feature and the
performance of both features in combination will be ranked against
other doublets. The combinatorial optimization algorithm proceeds
to then iteratively rank sets of three features (triplets). For
example, the first ranked doublet will be combined with another
feature to form a triplet of features, which will be ranked against
other triplets. The algorithm ranks singlets, doublets, triplets,
quadruplets and so on until the improvement in performance for
detecting a ROI reaches a predetermined value or point of
diminishing returns at which point a feature singlet, doublet or
triplet is extracted. For example, the highest ranked triplet may
provide only a small percent increase in performance over the
highest ranked doublet in which case, the doublet would be the
extracted feature combination to be employed in the next step in
classification. This estimation of the point of diminishing returns
advantageously enhances real-world performance by avoiding
data-overfitting or memorization during learning. The result of the
optimization algorithm is an object-level feature subset e.g. a
singlet, doublet, or triplet optimal feature selection also called
the optimal object-level feature subset (x). The optimal
object-level feature subset (x) is stored in the object-level data
file. The data file may be any of a number of local or global
databases.
[0129] The next step is the selection of the classification
algorithm 1706 in which an appropriate mapping function is selected
that transforms the optimal object-level feature subset (x) into a
discrete class label (y). The mapping function is implemented in
the form of a classifier. Classification algorithm selection 1706
is performed in the same way as in pixel and object processing.
Selection of the classification algorithm 1706 is performed
automatically or manually. The user may select a classification
algorithm manually from a GUI interface that provides a list of any
number of suitable algorithms. Typically, the selection involves a
choice between GMM and MVG. Alternatively, automatic selection of
the classification algorithm is data dependent and involves a
recommendation engine. The recommendation engine is an algorithm
that examines the object feature distribution and recommends the
MVG classifier for unimodal feature distributions and the GMM
classifier if the feature distribution is multimodal. If the actual
feature distribution looks multi-modal, GMM is the preferred
classifier and the number of clusters determines the number of
Gaussian mixtures to be used in representing each class conditional
feature distribution. If the actual feature distribution looks
unimodal, MVG is the preferred classifier. The recommendation
engine also estimates the most appropriate number of modes for the
GMM classifier.
[0130] As can be seen from above, there are several inputs for the
"OBJECT LEARN" processing option, which will now be summarized. The
inputs include the image file, which is typically a .jpg or .tif
file type, and the labeled image from pixel or subimage processing.
Another input is the ROI-membership array for identifying pixel
locations and their corresponding classification indices selected
manually by the user for all of the pixels in each object level
selection. Another input is a non-ROI or false alarm membership
array. Another input includes various processing parameters. These
processing parameters are stored in a 1.times.5 row vector wherein
the first parameter is the classifier selection. For example, if
the user manually inputs or if the recommendation engine
automatically selects MVG as the classifier of choice, the second
parameter will include a classifier identification such as "1" for
MVG and "2" for GMM. The second parameter is a classifier
parameter. This classifier parameter is dependent on the type of
classifier selected. For example, for the MVG classifier, the LLR
offset is the fourth parameter input where a positive LLR favors
ROI detection. For the GMM classifier, the fourth parameter input
is the number of Gaussian mixtures. The third and last element in
the processing parameters is a flag that is reserved for debugging
statements. For example, if this parameter is set to "1", then
debug statements will be written to a text file. Of course, another
input is a file name, which is used to concatenate object learn
data for particular tissue types for example. The file name input
includes both path and file name.
[0131] The output 1708 of "OBJECT LEARN" processing is a binary
object-level map for a two-class problem ("1" for non-ROI and "2"
for ROI) that is displayed on the computer monitor for user
inspection. The output also provides region labeling in which
connected pixels of the same class are aggregated into a region and
labeled as a ROI or non-ROI region. Each region is numbered and the
region label and number are stored in one or more databases. The
labeling can be performed in a variety of ways, and color is used
to denote a region identification number for easy user
identification of regions. FIG. 18 shows the results of object
processing.
[0132] The output of the "OBJECT LEARN" processing option is stored
in a local object-level database. The output of "OBJECT LEARN"
includes the original image, a further refined and labeled image, a
boundary structure summarizing classification results, and any
error message that may be applicable. The boundary structure that
summarizes classification results consists of several fields
including a list of pixel locations (x, y) and their corresponding
classification, an interior flag, the region area in pixels, the
length of the perimeter of a region and the number of boundary
pixels. The interior flag is set to one (1) if the boundary of a
region belongs to the interior of an object and set to zero (0)
otherwise. The list of pixel locations may be sorted, for example
by region and by classification.
[0133] Database creation and management 1710 is the same as that
discussed above with respect to pixel processing and described in
FIG. 9 with the word "pixel" being replaced with the word "object."
As shown in FIG. 4, object processing 410 is in electrical
communication with various local and global databases 414. As
additional ROI and non-ROI selections are made by the user when in
the "OBJECT LEARN" processing protocol, the optimal object-level
feature subset (x) is re-compiled for the entire data file either
local or global. Any local and global database to which object
processing results are concatenated also undergoes the
combinatorial optimization algorithm for a refinement of the
resident optimal object level feature subset.
[0134] The "OBJECT CLASSIFY" option will now be discussed. "OBJECT
CLASSIFY" employs data accumulated and stored in various databases
from one or more learning stages such as the "OBJECT LEARN" stage.
As discussed above, during "OBJECT LEARN" the user provides truth
annotations, for example ROI and non-ROI object-level selections
for representative images using GUI controls. Then, the system
discriminates ROIs from non-ROIs using classifiers. During the
learning stage, all the parameters and feature values associated
with pattern learning are stored for each image being investigated
in one or more database such as local and global databases. The
data from "OBJECT LEARN" is also concatenated to one or more global
databases as discussed above. During "OBJECT CLASSIFY" the system
utilizes the learned parameters stored in one or more databases to
perform automatic classification of regions as being ROIs or
non-ROIs. As the databases grow with time as more data is acquired
during "OBJECT LEARN," "OBJECT CLASSIFY" improves with age. The
"OBJECT CLASSIFY" operating mode is designed for high-throughput
batch processing with the human operator inspecting the processed
results for final decision.
[0135] Referring back to FIG. 16, after "OBJECT CLASSIFY" is
selected as the object processing option at 1604, "SUBIMAGE
CLASSIFY" will start 1608 by prompting the user via the GUI to load
an appropriate database 1622 if one is available. If no appropriate
database is available, the user will be directed to start "OBJECT
LEARN" 1606. A pull-down menu on the GUI provides a list of
appropriate local and global databases to be selected by the user
under "OBJECT CLASSIFY" and the user selects an object-level
database 1624. Typically, the available databases are ones created
during "OBJECT LEARN" and matched for the particular image at hand.
For example, if the image is a pancreatic tissue sample, the user
may select any local or global pancreatic tissue database created
at the object level. The next step is to run the classification
1626.
[0136] Running the classification under "OBJECT CLASSIFY" is
depicted in FIG. 19. After the start of "run classification" 1900,
the database that was selected at 1624 is accessed 1902 for its
critical algorithmic parameters and for the optimal object feature
subset. The critical algorithmic parameters are inputted along with
the image and labeled image from pixel or subimage processing. The
selected object-level database has already been updated at 1710 for
example. Hence, the database has undergone a ranking or re-ranking
of features and selection or re-selection of the optimal
object-level feature subset with each additional input of
information. Therefore, these values are ready for use with the
classification algorithm that is selected at 1904.
[0137] In "OBJECT CLASSIFY," the classification algorithm is
selected automatically. Automatic selection of the classification
algorithm is data dependent and involves a recommendation engine.
The recommendation engine is an algorithm that examines the feature
distribution and recommends the MVG classifier for unimodal
object-level feature distributions and the GMM classifier if the
object-level feature distributions are multimodal. The
recommendation engine also estimates the most appropriate number of
modes for the GMM classifier.
[0138] The output of "OBJECT CLASSIFY" is a binary image map for
the two-class problem ("1" for non-ROI and "2" for ROI) that is
displayed on the computer monitor for user inspection. The output
also provides region labeling in which connected pixels of the same
class are aggregated into a region and labeled as an object level
ROI or non-ROI region. Each region is numbered and the region label
and number are stored in one or more databases as shown in FIG. 18.
The labeling can be performed in a variety of ways, and color is
used to denote a region identification number for easy user
identification of regions.
[0139] The output 1906 of "OBJECT CLASSIFY" includes the original
image, the labeled image, a boundary structure summarizing
classification results, and any error message that may be
applicable. The boundary structure that summarizes classification
results consists of several fields including a list of pixel
locations (x, y), an interior flag, the region area in pixels, the
length of the perimeter of a region and the number of boundary
pixels. The interior flag is set to one (1) if the boundary of a
region belongs to the interior of an object and set to zero (0)
otherwise. The list of pixel locations may be sorted, for example
by region and by classification. The output is stored in a local or
global database.
[0140] Upon viewing the output of either the "OBJECT LEARN" or
"OBJECT CLASSIFY" options, the user determines that the output
binary image map is satisfactory and manually selects from the GUI
to proceed with LCM immediately or at a later time. Alternatively,
the processor may directly proceed with LCM without selection from
the GUI. If the tissue slide is loaded in the LCM device, the
processor will automatically position the laser at coordinates
corresponding to ROIs from the binary image map. Once positioned,
the laser is activated for LCM and the ROIs are automatically
transferred to the LCM transfer film and removed from the tissue
sample.
[0141] FIG. 20 shows the improving performance of tissue
recognition as the abstraction protocol proceeds from pixel to
subimage to object processing. After pixel processing, a lot of
unwanted regions above the epithelial layer still remain. However,
with subimage and object processing that utilize features at a
higher level of abstraction, almost all of the false detected
regions are filtered out, thereby greatly improving the accuracy of
extracting the right tissues using LCM equipment
[0142] While the present invention has been described with
reference to one or more particular variations, those skilled in
the art will recognize that many changes may be made thereto
without departing from the spirit and scope of the present
invention. Each of these embodiments and obvious various thereof
are contemplated as falling within the spirit and scope of the
claimed invention, which is set forth in the claims.
* * * * *
References