U.S. patent application number 10/437448 was filed with the patent office on 2004-01-22 for method and a system for combining automated psychiatric profiling from combined input images of brain scans with observed expert and automated interpreter using a neural network.
Invention is credited to Hillman, Yitzchak.
Application Number | 20040013291 10/437448 |
Document ID | / |
Family ID | 11074817 |
Filed Date | 2004-01-22 |
United States Patent
Application |
20040013291 |
Kind Code |
A1 |
Hillman, Yitzchak |
January 22, 2004 |
Method and a system for combining automated psychiatric profiling
from combined input images of brain scans with observed expert and
automated interpreter using a neural network
Abstract
A method for psychiatric profiling comprising: (a) obtaining a
3-D brain scan image (and its enhancement parameters) and obtaining
a psychiatric profile analysis; (b) extracting the edges of the
scan, pinpointing reference points on it, positioning,
standardizing, and aligning it; (c) autocropping, extracting a
plurality of features and/or regions within the scan; (d)
correlating said regions or features with database images and
parameters; (e) searching a message memory for messages that make
up an individual's profile (said messages correspond to feature/s
of a database), outputting each message to form a first profile set
of messages; (f) obtaining a second set of feature detections and
message statements; (g) accepting output detections and related
messages in the first set to form a third profile set (a subset of
the first set), combining the second and third profile to form a
fourth set, alternatively allowing the fourth set to equal the
first set, alternatively allowing the third set to equal the second
set; (h) storing in the message memory the fourth set,
corresponding to the brain scan image or storing the fourth set
which corresponds to new feature/s on the brain scan image,
providing a corrected output based on said corrected fourth set. A
system for providing human profiling using the method is also
disclosed.
Inventors: |
Hillman, Yitzchak;
(Jerusalem, IL) |
Correspondence
Address: |
LOWE HAUPTMAN GILMAN & BERNER, LLP
Suite 310
1700 Diagonal Road
Alexandria
VA
22314
US
|
Family ID: |
11074817 |
Appl. No.: |
10/437448 |
Filed: |
May 14, 2003 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
10437448 |
May 14, 2003 |
|
|
|
PCT/IL01/01047 |
Nov 12, 2001 |
|
|
|
Current U.S.
Class: |
382/128 ;
382/224 |
Current CPC
Class: |
G16H 30/20 20180101;
Y10S 128/923 20130101; Y10S 128/922 20130101; G16H 50/20
20180101 |
Class at
Publication: |
382/128 ;
382/224 |
International
Class: |
G06K 009/00; G06K
009/62 |
Foreign Application Data
Date |
Code |
Application Number |
Nov 14, 2000 |
IL |
139655 |
Claims
1. A method for providing human psychiatric profiling using a
process of analysis and classification of brain scan images
comprising the steps of; a) obtaining a 3-D brain scan image and
the result of a psychiatric profile analysis and parameters used to
enhance the image of the scan; b) extracting the edges of the brain
scan image, pinpointing reference points on it, positioning,
standardizing its size, and aligning it; c) autocropping,
extracting and masking a specified plurality of features and
regions and/or parameters within the brain scan; d) voting,
matching or correlating extracted regions, images and parameters of
a plurality of features of the scan with database template images
and parameters; e) searching in a message memory for a plurality of
messages that make up the profile of an individual, wherein each
message corresponds to the respective feature or combination of
features of a database, outputting each one of the said plurality
of messages concurrently to form a first profile set of messages;
f) obtaining a second set of feature detections and related message
statements; g) accepting some output detections and related
messages in the first set to form a third profile set of features
and related messages that is a subset of the first set, combining
the third profile set of messages with the second set to form a
fourth set, alternatively allowing the fourth set to equal the
first set, alternatively allowing the third set to equal the second
set; h) storing in the said message memory the fourth set of
detections and related messages corresponding to the said brain
scan image or storing in the said message memory the fourth set of
detections and related messages corresponding to a new combination
of features on the brain scan image, providing a corrected output
based on said corrected fourth set of detections and related
messages.
2. A method according to claim 1 wherein the said features of the
brain scan is one or a combination of general anatomic structures
including CSF, gray matter, ventricular fluid, and lesioned tissue
white matter, neurological mapping of activity to specified stimuli
(such as specific sight, sound, vocal, smell, touch, taste,
suggested imagination or other).
3. A method according to claim 1 wherein the said second set is
composed of none, one or a combination of the elements of the set
of feature detections and related message statements that form a
human profile made by an expert interpreter.
4. A method according to claim 1 wherein the said second set is
composed of none, one or a combination of the elements of the set
of feature detections and related message statements that form a
self profile of a person under analysis.
5. A method according to claim 1 where the input image is from an
MRI scanner, fMRI, MRS, PET, or other.
6. A method according to claim 1 in which the input image is
provided in a form of a computer memory of 2-D slices forming a 3-D
map or alternatively of a complete 3-D image.
7. A method according to claim 1 in which the pinpointing of
reference points is done by use of a matching template images.
8. A method according to claim 1 in which known reference points
are built into the input image.
9. A method according to claim 1 in which areas and features are
extracted using referencing to known given or calculated reference
points.
10. A method according to claim 1 wherein the psychiatric analysis
results are the profile results provided by readings of hand and
foot palms.
11. A method according to claim 1 in which a standardized
normalized image is determined using a generic algorithm that uses
the scanner image enhancement parameters as input parameters
provided into the generic algorithm procedure.
12. A method according to claims 3 and 4 wherein said detections
and related messages accepted from the first output set are
selected according to their likelihood of correct output detection
reporting and analysis.
13. A method according to claim 1 in which said edge extractor or
the position registration circuit, or the feature extractor,
comprises a neural network or in which the said pinpointing of
reference points on the brain scan is done after and as a result of
the said position registration using a neural network, or in which
the said voting, matching and correlating extracted regions and
images of features with database template images is done using a
neural network or in which the said storing of the fourth set of
detected features and related messages is in a form of a neural
network or in which detection is performed by brain scan detector
comprising a neural network.
14. A method according to claim 13 in which the said neural network
is a multi layer peceptron neural network.
15. A method according to claim 1 wherein the said pinpointing of
reference points is done by setting the palm, hand or foot in an
encompassing fixed shell before imaging thereby referencing from
the outer shell.
16. A method according to claim 1, additionally comprising a device
for measuring hardness and softness of specific mounts and areas of
the skin, the bending angle of the fingers and finger formations on
closed or clapped hands.
17. A method according to claims 1 and 16 wherein a mechanically
driven and controlled blunt pin element is used to press
automatically on the skin and palm mounts.
18. A method according to claims 1 and 16 wherein the pressure
applied is controlled and measured, rebound rate of the skin and
palm mount is measured using a laser scanner.
19. A method according to claim 1 in which auto-cropping and voting
is performed by a generic algorithm in which auto-cropping and
voting parameters are automatically optimized using a generic
algorithm that maximizes fitness.
20. A system for providing human profiling using the method as
defined in any of the preceding claims comprising of: a) A
mechanically driven blunt-pointed element adjoining an apparatus
for measuring the angle of finger bending; b) a mechanically driven
plate used for measuring the maximum allowed bending angle of the
finger adjoining the apparatus; c) RAM memory storage; d) an
microprocessor; e) input drive; f) a high resolution color printer;
g) a computer operating system; h) Internet and network connection.
Description
FIELD OF THE INVENTION
[0001] The present invention generally relates to a field of image
processing and data image classification. More particularly, the
present invention relates to a system and a method for detecting,
processing and classifying biometric images using digital
images.
BACKGROUND OF THE INVENTION
[0002] This invention uses a computer based technique to predict
brain disease, brain degenerative disease and atrophy as well as
other psychiatric illnesses before their onset. The brains of
people with Alzheimer show early atrophy before onset of diseased
symptoms. Schizophrenia patients show minor changes even before
their first psychotic episode. That raises the possibility of
screening and early diagnosis for the disease and early
intervention for people at risk.
[0003] This invention is an automated tool comprising a computed
algorithm for the sake of providing automated early diagnosis of
disease and psychiatric conditions.
[0004] There are many tools and procedures for obtaining brain scan
images. Likewise, there are countless algorithms and methods
intended to improve scan images using image processing techniques.
However, all automated diagnostic tools for brain scan images have
one thing in common. They must all contain within their algorithms
a method of data classification and storage as well as a method for
training the classifier using an expert interpreter. This invention
patents the use of a neural network (and as such includes a fuzzy
logic type of classifier).
[0005] Brain scan images are provided via an internet or network
connection and are analyzed by the procedures described.
[0006] Early treatment with behavioral therapy or drugs could
prevent, or at least mitigate, the full onset of Alzheimer or even
schizophrenia. The longer the disease or psychosis goes untreated,
the worse the outcome. Alzheimer and Schizophrenia is probably the
most expensive diseases for the National Health Service of any
country. If it can be prevented by early detection, the
implications are vast.
[0007] Magnetic resonance imaging (MRI) in brain scans showed
significant differences between healthy brains versus those of
patients. The brain changes began some time before the Alzheimer or
schizophrenic patients first suffered dementia or a psychotic
episode.
[0008] Over the clinical course of Alzheimer, patients demonstrate
progressive declines in functional ability that correlate with MMSE
scores. In the preclinical phase, also called MCI, patients with
MMSE score greater than 23 will demonstrate minimal
impairment--generally, mild memory loss--while functioning normally
and independently.
[0009] Though sensitivity issues are less of a problem in
diagnosing dementia per se, Specificity issues differentiating
Alzheimer from ordinary age related dementia proves a main hurdle.
MRI perfusion scan image with additional MRI structural imagery
proves to be an effective base image system to diagnose early
stages of Alzheimer using the Neural Network Computed method
described. Both Voxel-Based Morphometry and volumetric changes,
structural and functional variations are recorded on the database
for analysis using the neural network classifier. The advantages of
using a Neural network/Fuzzy logic type of analysis is that
structural atrophy can be classified not only by volumetric single
or small parameter system but by a multi parameter classifier of
normalized images having a multitude of variation of 3-D shapes in
a time dependent (age or durational progression of the disease)
axis. The spatial normalization step aims to map each structural
MRI to a template in standard 3-D and stereotactic space.
[0010] Atrophy rates for brain temporal lobe, cortex, Amygdalae,
temporal gyrus, hippocampus, and entorhinal cortices are
significantly increased in patients compared with controls. Linear
extrapolation backward suggested medial temporal lobe atrophy
commenced 3.5 years before onset of symptoms, when all patients
were asymptomatic. Medial temporal lobe atrophy rates are an early
and distinguishing feature of Alzheimer. Atrophy rates for brain,
temporal lobe, hippocampus, and entorhinal cortices are
significantly increased in patients compared with controls.
[0011] Schizophrenia patients have significant deficits in cortical
gray matter and in temporal lobe gray matter. The temporal lobes of
the brain are linked with speech and the experience of
hallucinations. There were also significant differences in whole
brain volume, as well as significant enlargement of the lateral and
third ventricles. Structural deviations were found in both
untreated and minimally treated subjects. No relationships were
found between any brain matter volumes and positive or negative
symptoms. Structural brain abnormalities were distributed
throughout the cortex with particular decrement evident in gray
matter. This feature is consistent with altered cell structure and
disturbed neuronal connectivity, which accounts for the functional
abnormality of psychosis. These brain abnormalities were not
specific to schizophrenia; they were also present in the brains of
people suffering from other kinds of psychosis, such as bipolar
disorder. It is assumed that many mental illnesses begin with the
same changes in brain structure and chemistry and that an initial
common pathway diverges into different forms of mental illness.
This means that treating anyone showing signs of the brain
abnormalities should prevent the onset of other mental diseases as
well.
[0012] Additionally, researchers in a pilot program at the Israel
"Nes-Ziona" psychiatric hospital found that it was possible to
determine psychiatric illnesses using the methods disclosed in
Israeli Patent Application No. 138975, especially emphasizing the
measurements of hardness of specific mounts and areas of the skin,
the bending angle of the fingers, spacing between the fingers,
relative finger lengths, the mounts on fingers, finger formations
on closed or clapped hands as well as other features of the palms
disclosed in Israeli Patent Application No. 138975.
[0013] The process of decoding and analyzing brain scan images so
as to provide an accurate psychiatric profile of individuals is
difficult to provide under human evaluation.
[0014] Therefore, the main objective of this invention is to
provide a method and system to diagnose and profile dementia
(especially Alzheimer) and psychiatric illness using images of
brain scans. Using MRI or other tools with brain scan analysis, the
present invention uses the creation of a neural network or a
multi-layer perceptron (MLP) neural network (NN) in which a
centralized data bank combines brain scan images with experience
from expert psychiatric advice and diagnosis placing emphasis on
medical and psychiatric history of individuals being analyzed. The
computer algorithms involved in this procedure have already proved
themselves clinically in other applications such as that described
in US patent Roger et al. (U.S. Pat. Nos. 6,205,236 and 5,999,639
and 6,115,488) where very similar Neural Network based algorithms
are currently used.
[0015] Evolutionary development of the human brain occurred at the
same time as the palms and during the first tool creation era of
the first humans. Human brain and palm morphologies resultantly
bear correlations. Therefore, another objective of the present
invention is to complement the above-mentioned method of diagnosis
and profiling with that disclosed in Israeli Patent Application No.
138975 whereby particular emphasis is made to certain features of
the hand and foot, mentioned above. These two objectives together
provide for more accurate psychiatric profiling. It is intended to
find correlations between palm hand and foot features, and features
on brain scans. This may provide insight into psychiatric,
psychological and character profiling. This is important in brain
research as well as in providing more accurate diagnostics.
[0016] Another objective of the present invention is the
classification of brain scans using MRS and fMRI (Magnetic
resonance Spectroscopy and Functional MRI) indicating functional
characteristics of the brain (i.e neural activity).
[0017] It is assumed that different classifications of character,
personality, psychological and psychiatric profiles would have a
different spread of neural activity for similar neural stimuli,
such as specific sight, sound, vocal, smell, touch, taste,
suggested imagination or other. It is intended to find and use
unique specific neural activity associated with each of these
specified classifications indicating the link between the neural
activity and the classification. Finding such a link and
classifying it in the form of a computed neural network will aid in
the psychiatric diagnosis, making it more accurate.
[0018] Using brain scan technology, we are now able to identify the
content of a person's thought, albeit in a very limited context.
However, it is assumed that although, the basic pattern of neural
firing is maintained in the general population, significant
variations on the general pattern apply. These variations are
dependant amongst factors that include the psychiatric profile of
the person.
[0019] In many previous studies have shown that brain areas can be
selective for processing a particular type of visual information.
In the cortical brain regions associated with mental processing,
the fusiform face area responds strongly to faces while the
para-hippocampus place area responds strongly to indoor and outdoor
scenes depicting the layout of local space. It was also found that
the magnitude of activity in these two brain areas is much livelier
or stronger when one is seeing the picture (physically present in
front of them) compared with just imagining it.
[0020] Portable scanning technique (such as laser scanners) could
be used to gain some insight into what is happening in the minds of
people who are unable to communicate because they are suffering
from an injury or disorder that makes speech impossible. However,
it is assumed that it will be possible to predict and analyze
thought patterns with almost 100% accuracy if adjustment is made
for the thought pattern analysis by taking into consideration the
psychiatric profile of the individual being analyzed. Therefore,
another objective of this patent is to categorize neural functional
activity (agitated by specified stimuli) according to the
psychiatric profile thereby providing for a method and system for
analyzing thoughts. This procedure has special emphasis for the
need of prostheses limbs in order to function.
[0021] A computed neural network is used to correlate sequenced
brain neural activity with memorized sequences of template scan
images recorded in a central database of template scan images that
have been classified according to their psychiatric profile.
[0022] Other objectives and advantages of the invention will be
apparent from the following detailed description that follows.
[0023] In the present invention, the terms "psychiatric profiling"
or "diagnosis" are intended to include profiling such as medical,
psychiatric, genetic, psychological and character profiling.
SUMMARY OF THE INVENTION
[0024] There is thus provided in the present invention a method for
providing human psychiatric profiling using a process of analysis
and classification of brain scan images comprising the steps of; a)
obtaining a 3-D brain scan image and the result of a psychiatric
profile analysis and parameters used to enhance the image of the
scan; b) extracting the edges of the brain scan image, pinpointing
reference points on it, positioning, standardizing its size, and
aligning it; c) autocropping and extracting a specified plurality
of features and regions and/or parameters within the brain scan; d)
voting, matching or correlating extracted regions, images and
parameters of a plurality of features of the scan with database
template images and parameters; e) searching in a message memory
for a plurality of messages that make up the profile of an
individual, wherein each message corresponds to the respective
feature or combination of features of a database, outputting each
one of the said plurality of messages concurrently to form a first
profile set of messages; f) obtaining a second set of feature
detections and related message statements; g) accepting some output
detections and related messages in the first set to form a third
profile set of features and related messages that is a subset of
the first set, combining the third profile set of messages with the
second set to form a fourth set alternatively allowing the fourth
set to equal the first set, alternatively allowing the third set to
equal the second set; h) storing in the said message memory the
fourth set of detections and related messages corresponding to the
said brain scan image or storing in the said message memory the
fourth set of detections and related messages corresponding to a
new combination of features on the brain scan image, providing a
corrected output based on said corrected fourth set of detections
and related messages.
[0025] According to one preferred embodiment of the method, the
said features of the brain scan is one or a combination of general
anatomic structures including CSF, gray matter, ventricular fluid,
and lesioned tissue white matter, neurological mapping of activity
to specified stimuli (such as specific sight, sound, vocal, smell,
touch, taste, suggested imagination or other).
[0026] According to a preferred embodiment of the method, the said
second set is composed of none, one or a combination of the
elements of the set of feature detections and related message
statements that form a human profile made by an expert
interpreter.
[0027] In one embodiment, said second set is composed of none, one
or a combination of the elements of the set of feature detections
and related message statements that form a self profile of a person
under analysis. In such case, in one embodiment the detections and
related messages accepted from the first output set are selected
according to their likelihood of correct output detection reporting
and analysis.
[0028] In another embodiment, the input image is from an MRI
scanner, fMRI, MRS, PET, CAT, SPECT, EEG, laser, or other.
[0029] In another embodiment the input image is provided in a form
of a computer memory of 2-D slices forming a 3-D map or
alternatively of a complete 3-D image.
[0030] In another embodiment the pinpointing of reference points is
done by use of a matching template images.
[0031] In another embodiment, known reference points are built into
the input image.
[0032] In another embodiment areas and features are extracted using
referencing to known given or calculated reference points.
[0033] In another embodiment the psychiatric analysis results are
the profile results provided by readings of hand and foot
palms.
[0034] In another embodiment a standardized normalized image is
determined using a generic algorithm that uses the scanner image
enhancement parameters as input parameters provided into the
generic algorithm procedure.
[0035] In another embodiment said edge extractor or the position
registration circuit, or the feature extractor, comprises a neural
network or in which the said pinpointing of reference points on the
brain scan is done after and as a result of the said position
registration using a neural network, or in which the said voting,
matching and correlating extracted regions and images of features
with database template images is done using a neural network or in
which the said storing of the fourth set of detected features and
related messages is in a form of a neural network or in which
detection is performed by brain scan detector comprising a neural
network.
[0036] In such case, In one embodiment the said neural network is a
multi layer peceptron neural network.
[0037] In another embodiment the said pinpointing of reference
points is done by setting the palm, hand or foot in an encompassing
fixed shell before imaging thereby referencing from the outer
shell.
[0038] In one embodiment the method is additionally comprising a
device for measuring hardness and softness of specific mounts and
areas of the skin, the bending angle of the fingers and finger
formations on closed or clapped hands. In such case, in one
embodiment, a mechanically driven and controlled blunt pin element
is used to press automatically on the skin and palm mounts. In
another embodiment the pressure applied is controlled and measured,
rebound rate of the skin and palm mount is measured using a laser
scanner.
[0039] In one embodiment auto-cropping and voting are performed by
a generic algorithm in which auto-cropping and voting parameters
are automatically optimized using a generic algorithm that
maximizes fitness.
[0040] There is also provided in the present invention a system for
providing human profiling using the method as defined in any of the
preceding claims comprising of: a) A mechanically driven
blunt-pointed element adjoining an apparatus for measuring the
angle of finger bending; b) a mechanically driven plate used for
measuring the maximum allowed bending angle of the finger adjoining
the apparatus; c) RAM memory storage; d) an microprocessor; e)
input drive; f) a high resolution color printer; g) a computer
operating system.
DETAILED DESCRIPTION OF THE INVENTION
[0041] A brain scan is provided using conventional brain scanning
techniques.
[0042] Parameters used in obtaining the scan are provided. These
parameters indicate either filtering, thresholding or other image
enhancing parameters used in obtaining the scanned image. Brain
scans of different "slices" and plains at differing given angles of
the brain make up the input image to the system providing for a 3-D
image of the brain. This scan is stored in memory. Ordinary MRI may
map gray and white matter, ventricular fluid, and lesioned tissue
using both or either T1 or T2 times. MRS fMRI and PET scans give
other mappings.
[0043] In order to normalise and standardize the scans, into a
standard scan image, a generic algorithm is used. Scan image
normalisation uses the input parameters provided with the original
brain scan image as parameters used in this generic algorithm.
[0044] A feature extractor is used for finding reference points on
the brain image.
[0045] Pinpointing reference points is done automatically by
matching template images of the brain to database images of brains.
A second feature extractor process or circuit is provided for
extracting all the features necessary for profiling analysis of an
individual. These include specific 2-D slices or plains on the 3-D
brain scan image at specific brain areas and angles. Areas and
features of these images are extracted by using a process of
referencing from a given set of reference points on similar brain
scan images.
[0046] A protocol for brain extraction and automatic tissue
segmentation of MR images involves the brain extraction algorithm,
proton density and T2-weighted images used to generate a brain mask
encompassing the full intracranial cavity. Segmentation of brain
tissues into gray matter (GM), white matter (WM), and cerebral
spinal fluid (CSF) is accomplished on a T1-weighted image after
applying the brain mask. The fully automatic segmentation algorithm
is histogram-based and uses the Expectation Maximization algorithm
to model a four-Gaussian mixture for both global and local
histograms. The means of the local Gaussians for GM, WM, and CSF
are used to set local thresholds for tissue classification.
Reproducibility at the regional level by comparing segmentation
results within the 12 major Talairach subdivisions.
[0047] A voting process or circuit compares the extracted brain
scan features with a database of previously extracted brain scan
features to categorize the object within a set of objects having
similar or highly correlated images of the features by use of a
neural network.
[0048] The results of the system are optimally combined with the
results given by the neural network computation.
[0049] Additional measurements of palm hand or foot are made. In
order to measure hardness and softness of the palms of hand and
foot regions, specific regions on the hand and foot are pressed
using a mechanically driven and controlled blunt pin element that
is pressed automatically on the skin and palm mounts. The pressure
applied is controlled and measured. Rebound rate of the skin and
palm mount is measured using the laser scanner as listed in patent
Israeli Patent Application No. 138975.
[0050] Similarly, in order to measure the maximum bending angles of
the fingers, automated controlled and measured pressure is applied
on the fingers using a mechanically driven plate while measuring
the maximum allowable bending angle of tie finger. An edge
extractor processes the brain scan images in order to determine the
edges of the brain in the image. This is done simply by matching
template images of objects having pre-determined outer edges
declared as belonging to the object features.
[0051] Auto-cropping is performed by one of many methods. Auto
cropping of specific regions on the brain scan images is optimized
by parameter-optimizing means using a genetic algorithm (GA) so as
to maximize the true-positive image detection rate while minimizing
the false-positive detection rate. Of course, other optimization
schemes may be used as well. Preferably, the cropping is performed
automatically, although the images could be cropped manually, and
the results stored as potential templates used for additional
automatic classification.
[0052] Generic algorithms search the solution space to maximize a
fitness (objective) function by use of simulated evolutionary
operators. In the present invention, the fitness function to be
maximized reflects the goals of maximizing the number of
true-positive pixel elements of major lines while minimizing the
number of false-positive detections. The use of generic algorithms
requires determination of several issues: objective function
design, parameter set representation, population initialization,
choice of selection function, choice of genetic operators
(reproduction mechanisms) for simulated evolution, and
identification of termination criteria.
[0053] The design of the objective function is a key factor in the
performance of any optimization algorithm. The function
optimization problem for detecting brain scan image features may be
described as follows: given some finite domain, D, a particular set
of feature detection parameters, x={t, f, k. sub. lo, k. sub. hl, .
. . , d} where x is an element of D, and an objective function f.
sub. obj. where x denotes the set of real numbers, find the x in D
that maximizes or minimizes f. sub. obj. Optimization may be
achieved by maximizing the true positive rate (TP) for a feature
relating to a given profile assessment message subject to the
constraint of minimizing the false positive (FP) rate. Assuming TN
represents profile elements and features correctly identified as
not belonging to our objects and FP represents profile elements and
features reported as belonging to our objects under investigation.
TP is the set of profile elements and features reported by a CAD,
and FN is set of profile elements and features that are known to be
true and that are not reported by CAD.
[0054] It is assumed systems may be optimized to maximize the TP
and additional FN rates subject to the constraint of minimizing the
FP rate. Different objective functions may be used.
[0055] A real-valued GA is an order of magnitude more efficient in
CPU time than the binary GA, and provides higher precision with
more consistent results across replications.
[0056] For that reason, this embodiment of the present invention
uses a floating-point representation of the generic algorithm.
[0057] This embodiment also seeds the initial population with some
members known beforehand to be in an interesting part of the search
space so as to iteratively improve existing solutions. Also, the
number of members is limited to twenty or some other pre-determined
number so as to reduce the computational cost of evaluating
objective functions.
[0058] In one embodiment of the invention, normalized geometric
ranking is used, as discussed in greater detail in Houck, et al.,
supra, for the probabilistic selection process used to identify
candidates for reproduction. Ranking is less prone to premature
convergence caused by individuals that are far above average. The
basic idea of ranking is to select solutions for the mating pool
based on the relative fitness between solutions. This embodiment
also uses the default genetic operation schemes of arithmetic
crossover and non-uniform mutation included in Houck, et al.'s
GA.
[0059] This embodiment continues to search for solutions until the
objective function converges. Alternatively, the search could be
terminated after a predetermined number of generations. Although
termination due to loss of population diversity and/or lack of
improvement is efficient when crossover is the primary source of
variation in a population, homogeneous populations can be succeeded
with better (higher) fitness when using mutation. Crossover refers
to generating new members of a population by combining elements
from several of the most fitting members. This corresponds to
keeping solutions in the best part of the search space. Mutation
refers to randomly altering elements from the most fitting members.
This allows the algorithm to exit an area of the search space that
may be just a local maximum. Since restarting populations that may
have converged proves useful, several iterations of the GA are run
until a consistent lack of increase in average fitness is
recognized.
[0060] Once potentially optimum solutions are found by using the
GA, the most fitting GA solution may be further optimized by local
searches. An alternative embodiment of the invention uses the
simplex method to further refine the optimized GA solution.
[0061] The auto-cropping system may also benefit from optimization
of its parameters including contrast value, number of erodes,
number of dilates and other parameters.
[0062] The method for optimizing the auto-cropper includes the
steps of generating line masks by hand for some training data,
selecting an initial population, and producing line masks for
training data. The method further includes the steps of measuring
the percent of overlap of the hand-generated and automatically
generated masks as well as the fraction of auto-cropped features
outside the hand-generated masks. The method further comprises
selecting winning members, generating new members, and iterating in
a like manner as described above until a predetermined objective
function converges.
[0063] Thresholding, contrast and image enhancing parameters used
by a particular brain scanner may be assumed as input parameters
that are fed into system and associated with the particular brain
scan image. These parameters are used for standardizing and
normalizing the scanned image using generic algorithm
techniques.
[0064] Feature extraction is obtained by first identifying and
aligning the image brain scan using template matching then by use
of further template matching, a point on the object is chosen as a
reference point. Features are then extracted by template matching
with reference to the different reference points such that the
bigger the brain area size, the larger the area chosen for template
matching. This brain size image adjustment is controlled by a
parameter that is included amongst the optimization parameters
optimized in the feature detection and auto cropping process.
[0065] Relevant features within objects are obtained according to
the invention by providing a novel method and system for automated
feature detection from digital object images. Parameters necessary
for cropping the relevant digital feature images are optimized; the
digital feature images are cropped based on the optimized cropping
parameters for selecting profile and relevant feature for further
analysis.
[0066] The detected features and relating profiles are then stored
as a detection image and profile, the detection image and profile
is processed for display, and a computer-aided detection image is
produced for review by an expert such as a psychiatrist etc.
[0067] The expert first reviews the original scan image, reports a
profile and a set of suspicious regions and features of interest
that diagnose the particular profile and feature set, S1. S1 is a
subset of all possible profiles and features S of the objects under
investigation, A CAD (computer aided diagnosis) system, or more
particularly, the CAD system of the invention, operates on the
original set of suspicious regions and features and reports a
second set of suspicious diagnosis or regions of interest, which
form profile and features set S2. The expert then re-examines the
set S2, accepts, or rejects members of set S2, thus forming a third
profile set S3 that is a subset of set S2. The expert then forms
another set S4 that is a set of all profile attributes that belong
to S1 in union with profile attributes S3. The workup regions in S4
and the patients under analysis having S4 are then recommended for
further psychiatric examination and diagnosis.
[0068] CAD system outputs are thereby incorporated with the
expert's analysis in a way that optimizes the overall sensitivity
of detecting true positive features and regions of interest as well
as associated profile assessments.
[0069] The digital images are stored as digital representations of
the original feature images on computer-readable storage media. In
a preferred embodiment, the digital representations or images are
stored on a 12 GB hard drive of a general-purpose computer such as
a PC having dual Pentium III microprocessors running at 566 MHZ,
512 MB of RAM memory, a high resolution color monitor, a pointing
device, and a high resolution color inkjet HP printer. The system
operates within a Windows 2000 operating system connected via a
modem to the Internet so as to receive and send results from around
the globe via a worldwide network.
[0070] Template features are provided as inputs to the classifier,
which classifies each template or combinations of templates as
being associated with particular psychiatric or psychological set
of profile elements "statements".
[0071] In practice, a feature detector is only able to locate
regions of interest in the digital representation of the original
object that may be associated with a particular profile element or
"statement". In any detector, there is a tradeoff between locating
as many potentially suspicious regions as possible versus reducing
the number of normal regions falsely detected as being potentially
suspicious. CAD systems are designed to provide the largest feature
detection rates possible at the expense of detecting potentially
significant numbers of irrelevant regions. Many of these unwanted
detections are removed from consideration by applying pattern
recognition techniques.
[0072] Pattern recognition is the process of making decisions based
on measurements. In this system, regions of interest or detections
are located by a detector, and then accepted or rejected for
display. The first step in the process is to characterize the
detected regions. Toward this end, multiple measurements are
computed from each of the detected regions. Each measurement is
referred to as a feature. A collection of measurements for a
detected region is referred to as a feature vector, wherein each
element of the vector represents a feature value. The feature
vector is input to a discriminant function. A classifier has a
feature vector x applied to a set of discriminant functions g (x).
A discriminant function computes a single value as a function of an
input feature vector. Discriminant functions may be learned from
training data and implemented in a variety of functional forms. The
output of a discriminant function is referred to as a test
statistic. Classification is selecting a class according to the
discriminant function with the greatest output value. The test
statistic is compared to a threshold value. For values of the test
statistic above the threshold, the profile set associated with the
feature vector is retained and displayed as potentially suspicious.
When the test statistic is below the threshold, the profile set is
not displayed.
[0073] Many methods are available for designing discriminate
functions. One approach considered for this invention is a class of
artificial neural networks. Artificial neural networks require
training, whereby the discriminate function is formed with the
assistance of labeled training data. In a preferred embodiment, the
classification process is implemented by means of a multi-layer
perceptron (MLP) neural network (NN). Of course, other classifier
means could be used such as, for example, a statistical quadratic
classifier.
[0074] The embodiment of the MLP NN system is implemented by means
of software running on a general-purpose computer. Alternatively,
the MLP NN could also be implemented in a hardware configuration by
means readily obtained apparent to those with ordinary skill in the
art.
[0075] The weight values are obtained by training the network.
Training consists of repeatedly presenting feature vectors of known
class membership as inputs to the network. Weight values are
adjusted with a back propagation algorithm to reduce the mean
squared error between actual and desired network outputs. Desired
outputs of z. sub. 1 and z. sub. 2 for a suspicious input are +1
and -1, respectively. Desired outputs of z. sub. 1 and z. sub. 2
for non-suspicious inputs are -1 and +1, respectively. Other error
metrics and output values may also be used.
* * * * *