U.S. patent application number 14/147985 was filed with the patent office on 2014-07-10 for reconstruction method of biological tissue image, apparatus therefor, and image display apparatus using the biological tissue image.
This patent application is currently assigned to CANON KABUSHIKI KAISHA. The applicant listed for this patent is CANON KABUSHIKI KAISHA. Invention is credited to Koichi Tanji.
Application Number | 20140193060 14/147985 |
Document ID | / |
Family ID | 51061001 |
Filed Date | 2014-07-10 |
United States Patent
Application |
20140193060 |
Kind Code |
A1 |
Tanji; Koichi |
July 10, 2014 |
RECONSTRUCTION METHOD OF BIOLOGICAL TISSUE IMAGE, APPARATUS
THEREFOR, AND IMAGE DISPLAY APPARATUS USING THE BIOLOGICAL TISSUE
IMAGE
Abstract
The present invention provides a method for classifying
biological tissues with high precision compared to a conventional
method. When measuring a spectrum which has a two-dimensional
distribution that is correlated with a slice of a biological
tissue, and acquiring a biological tissue image from the
two-dimensional measured spectrum, the method includes dividing an
image region into a plurality of small blocks, and then
reconstructing the biological tissue image by using the measured
spectrum and a classifier corresponding to each of the regions.
Inventors: |
Tanji; Koichi;
(Kawasaki-shi, JP) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
CANON KABUSHIKI KAISHA |
Tokyo |
|
JP |
|
|
Assignee: |
CANON KABUSHIKI KAISHA
Tokyo
JP
|
Family ID: |
51061001 |
Appl. No.: |
14/147985 |
Filed: |
January 6, 2014 |
Current U.S.
Class: |
382/133 |
Current CPC
Class: |
G06T 7/11 20170101; G06T
11/003 20130101; G06T 2207/30024 20130101; G06T 2207/20081
20130101; G06T 2207/20021 20130101; G06K 9/00147 20130101; G06T
7/0014 20130101; G06T 2207/10056 20130101 |
Class at
Publication: |
382/133 |
International
Class: |
G06K 9/00 20060101
G06K009/00; G06T 7/00 20060101 G06T007/00; G06T 11/00 20060101
G06T011/00 |
Foreign Application Data
Date |
Code |
Application Number |
Jan 8, 2013 |
JP |
2013-000883 |
Aug 6, 2013 |
JP |
2013-163399 |
Dec 4, 2013 |
JP |
2013-251050 |
Claims
1. A reconstruction method of a biological tissue image using a
signal processing apparatus based on a measured spectrum correlated
with a substance distributed within a biological tissue comprising:
acquiring, within an image region, the measured spectrum at each of
plural points within the biological tissue; dividing the image
region into a plurality of small blocks; selecting one or more of
peaks of the measured spectrum in each of the small blocks;
acquiring a classifier corresponding to each of the small blocks;
and acquiring the biological tissue image per each of the small
blocks based on the selected one or more of peaks and the
corresponding classifier.
2. The reconstruction method of the biological tissue image
according to claim 1, wherein the classifier to be applied to an
image region is generated from classification conditions acquired
per each of a plurality of small blocks, by a regression analysis
of the classification conditions, and then the biological tissue
image in the image region is generated by applying the classifier
to the measured spectrum.
3. The reconstruction method of the biological tissue image
according to claim 1, wherein the biological tissue images acquired
per each of a plurality of small blocks are integrated to generate
the biological tissue image in the image region.
4. The reconstruction method of the biological tissue image
according to claim 1, wherein the classifier is generated by
applying a training data to the measured spectrum.
5. The reconstruction method of the biological tissue image
according to claim 1, wherein, at the selecting one or more of
peaks of the measured spectrum, the peak for use in the
classification is determined based on Mahalanobis distance defined
by a ratio of an inter-group dispersion to an intra-group
dispersion.
6. The reconstruction method of the biological tissue image
according to claim 1, wherein, the measured spectrum is any of a
spectrum in an ultraviolet region, a visible region and an infrared
region, a Raman spectroscopic spectrum, and a mass spectrum.
7. The reconstruction method of the biological tissue image
according to claim 1, wherein, the biological tissue is a
pathological tissue.
8. A biological tissue image acquiring apparatus, wherein a
biological tissue image is reconstructed by the method according to
claim 1.
9. An image display apparatus, wherein, at a pathological
diagnosis, a lesion is displayed by the biological tissue image
acquiring apparatus according to claim 7.
10. A reconstruction method of a biological tissue image using a
signal processing apparatus based on a measured spectrum correlated
with a substance distributed within a biological tissue comprising:
acquiring the measured spectrum within the biological tissue;
acquiring morphological information from distribution information
of a peak component in the measured spectrum; acquiring a
classifier; and applying the classifier to both of the measured
spectrum of the biological tissue and of the morphological
information acquired from the distribution information of the peak
component in the measured spectrum, to acquire the biological
tissue image.
11. The reconstruction method of the biological tissue image
according to claim 10, wherein when the classifier is generated and
the biological tissue image is reconstructed, both of the measured
spectrum and a higher-order local autocorrelation which is acquired
from distribution information thereof are used.
Description
BACKGROUND OF THE INVENTION
[0001] 1. Field of the Invention
[0002] The present invention relates to a reconstruction method of
a biological tissue image and an apparatus therefor, and
particularly relates to a method for reconstructing a biological
tissue image from measured spectrum data which is correlated with a
substance distributed within a biological tissue, and to an
apparatus therefor. The present invention also relates to an image
display apparatus for clearly displaying a lesion at a pathological
diagnosis by using thus acquired biological tissue image.
[0003] 2. Description of the Related Art
[0004] Conventionally, a pathological diagnosis has been conducted
which is specifically a diagnosis for the presence or absence of a
lesion and a type of the lesion, based on the observation for a
biological tissue of an object by a microscope or the like. In the
pathological diagnosis, a constitutive substance and a contained
substance which are correlated with a biological tissue of an
object to be observed are required to be visualized. So far, a
technique for staining a specific antigen protein by using an
immunostaining method has mainly been employed in the pathological
diagnosis. When breast cancer is taken as an example, ER (estrogen
receptor which is expressed in hormone-dependent tumor) which
serves as a determination criterion for hormone therapy and HER2
(membrane protein to be found in fast-growing cancer) which serves
as a determination criterion for Herceptin administration are
visualized by the immunostaining method. However, the
immunostaining method has such problems that the reproducibility is
poor because an antibody is unstable and antigen-antibody reaction
efficiency is difficult to be controlled. In addition, when needs
of such a functional diagnosis will be grown in the future, for
instance, and when there arises a need of detecting several tens or
more types of constitutive substances or contained substances, the
currently-employed immunostaining method has a problem of being
incapable of meeting the need any more.
[0005] In addition, in some cases, the visualization of the
substance which is distributed within a biological tissue, such as
the constitutive substance and the contained substance, is not
sufficient at a tissue level, and the visualization at a cellular
level is required. For instance, in research on a cancer stem cell,
it was revealed that a tumor was formed in only part of fractions
of a tumor tissue after xenotransplantation to immunocompromised
mice, and accordingly, it is being understood that the growth of a
tumor tissue is dependent on differentiation and self-reproduction
abilities of the cancer stem cells. In such examination, it is
necessary not to observe the entire tissue, but to observe an
expression distribution of a constitutive substance or a contained
substance in each of individual cells in a tissue.
[0006] Incidentally, the above described "cellular level" means a
level at which at least each of the individual cells can be
classified. A diameter of the cell exists in a range of
approximately 10 .mu.m to 20 .mu.m (provided that large cell such
as nerve cell has diameter of about 50 .mu.m). Accordingly, in
order to acquire a two-dimensional distribution image at a cellular
level, the spatial resolution needs to be 10 .mu.m or less, can be
5 .mu.m or less, further can be 2 .mu.m or less, and still further
can be 1 .mu.m or less. The spatial resolution can be determined
from a result of, for instance, a linear analysis of a knife-edge
specimen. In other words, the spatial resolution is determined
based on the general definition of "a distance between two points
at which signal intensities originating in a concerned substance in
the vicinity of the boundary of a specimen are 20% and 80%,
respectively."
[0007] As described above, in the pathological diagnosis, the
constitutive substance and the contained substance which are
correlated with a lesion or a pathological tissue are required to
be exhaustively visualized at a cellular level. The lesion or the
pathological tissue means, for instance, a tumor tissue and the
like. Candidates for a method of such visualization include
secondary-ion mass spectrometry (SIMS) including time-of-flight
secondary-ion mass spectrometry (TOF-SIMS). A mass spectrum is used
as a measured spectrum. Furthermore, the candidates include also
Raman spectroscopy. Usable measured spectra include spectra in an
ultraviolet region, a visible region and an infrared region. These
measurement methods can provide information at each of plural
points in a space at high spatial resolution. Specifically, the
measurement methods can provide spatial distribution information on
each peak value of the measured spectrum which is correlated with a
substance that is an object to be measured, and accordingly, can
determine a spatial distribution of the substance in a biological
tissue which is correlated with the measured spectrum.
[0008] An SIMS method is a method of obtaining a mass spectrum at
each point on a specimen by irradiating the specimen with a primary
ion beam and detecting secondary ions which have been separated
from the specimen. In a TOF-SIMS, for instance, it is possible to
obtain the mass spectrum at each point on the specimen by
identifying the secondary ion with the use of such a fact that a
flight time of the secondary ion depends on a mass m and an
electric charge z of the ion.
[0009] A Raman spectroscopy acquires a Raman spectrum by
irradiating a substance with a laser beam which is a monochromatic
light as a light source, and detecting the generated Raman
scattering light with a spectroscope or an interferometer. A
difference (Raman shift) between a frequency of the Raman
scattering light and a frequency of incident light takes a value
peculiar to the structure of the substance, and accordingly the
Raman spectroscopy can acquire the Raman spectrum peculiar to an
object to be measured.
[0010] In order to acquire biological information from data of the
measured spectrum, a conventional method has generated a classifier
beforehand by machine learning, and has applied the generated
classifier to the data of the measured spectrum of the specimen
(see Japanese Patent Application Laid-Open No. 2010-71953). On the
other hand, it has been attempted to overlap a measured spectrum
image (spectrum information) with an optical image (morphological
information) and display the overlapped image, because a biological
tissue image is indispensable in a pathological diagnosis (see
Japanese Patent Application Laid-Open No. 2010-85219).
Incidentally, the machine learning described here means a technique
of empirically learning data which have been previously acquired,
and interpreting newly acquired data based on the learning results.
Further, the classifier refers to determination criterion
information to be generated by empirically learning a relationship
between previously acquired data and biological information.
[0011] Conventionally, an example of diagnosing a disease by
applying the classifier which has been generated by the machine
learning is described also in Patent Document 1. The object to be
diagnosed is one measured spectrum data (for one point on space or
whole specimen), and it has not been assumed to acquire the
biological tissue image from a spatial distribution of the measured
spectrum. In addition, there is an example of overlapping the
measured spectrum image (spectrum information) with the optical
image (morphological information), but there has been no example of
acquiring the biological tissue image by applying the machine
learning (classifier) to both the spectrum information and the
morphological information. Specifically, such a method has not been
disclosed as to reconstruct a biological tissue image with high
precision, which displays a diagnosis result related to a presence
or an absence of a cancer and the like, from a result of having
measured a spectrum having the spatial distribution for the
biological tissue of an object.
[0012] In addition, when the measured spectrum has the spatial
distribution, the characteristics of the data are different between
positions at which the data is measured, for instance, a datum
measured in the middle part of the image is different in the
characteristics from that measured in the peripheral portion of the
image. Accordingly, the classifier suitable for the position needs
to be applied according to the position at which the data is
measured. However, conventionally, such a method has not been
disclosed as to have assumed such a situation.
SUMMARY OF THE INVENTION
[0013] According to one aspect of the present invention, there is
provided a reconstruction method of a biological tissue image using
a signal processing apparatus based on a measured spectrum
correlated with a substance distributed within a biological tissue,
which includes: acquiring, within an image region, the measured
spectrum at each of plural points within the biological tissue;
dividing the image region into a plurality of small blocks;
selecting one or more of peaks of the measured spectrum in each of
the small blocks; acquiring a classifier corresponding to each of
the small blocks; and acquiring the biological tissue image per
each of the small blocks based on the corresponding classifier.
[0014] According to another aspect of the present invention, there
is provided a reconstruction method of a biological tissue image
using a signal processing apparatus based on a measured spectrum
correlated with a substance distributed within a biological tissue,
which includes: acquiring the measured spectrum of the biological
tissue; acquiring morphological information acquired from
distribution information of a peak component in the measured
spectrum; acquiring a classifier; and applying the classifier to
both of the measured spectrum of the biological tissue and of the
morphological information acquired from the distribution
information of the peak component in the measured spectrum, to
acquire the biological tissue image.
[0015] The biological tissue image acquired by the method of the
present invention can be used for pathological diagnosis and the
like.
[0016] Further features of the present invention will become
apparent from the following description of exemplary embodiments
with reference to the attached drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
[0017] FIG. 1 is a schematic view of an apparatus on which the
present invention is mounted.
[0018] FIG. 2 is a schematic view of a spectrum signal having an
intensity distribution in a two-dimensional plane.
[0019] FIGS. 3A, 3B and 3C are conceptual views of peak components
in a spectrum.
[0020] FIG. 4 is a flow chart of the present invention.
[0021] FIG. 5 is a flow chart of machine learning with the use of a
classification analysis of a block of the present invention.
[0022] FIGS. 6A, 6B and 6C are schematic views of the discriminant
analysis of an image block.
[0023] FIGS. 7A, 7B and 7C are views schematically illustrating a
projection axis which maximizes a ratio of an inter-group
dispersion to an intra-group dispersion.
[0024] FIGS. 8A, 8B and 8C are views schematically illustrating a
state in which confounding occurs due to data with different
conditions.
[0025] FIGS. 9A, 9B and 9C are schematic views illustrating a
process of determining a regression model by the discriminant
analysis of the block.
[0026] FIGS. 10A, 10B and 10C are views schematically illustrating
a series of processes of the present invention.
[0027] FIGS. 11A, 11B and 11C are views illustrating an application
process of a first exemplary embodiment of the present
invention.
[0028] FIGS. 12A and 12B are views illustrating that classification
conditions are different among the blocks of the different
images.
[0029] FIGS. 13A and 13B are views illustrating that confounding
occurs in data.
[0030] FIGS. 14A and 14B are views illustrating an application
result of the regression analysis in the first exemplary embodiment
of the present invention.
[0031] FIGS. 15A and 15B are views illustrating an application
effect of the first exemplary embodiment of the present
invention.
[0032] FIGS. 16A and 16B are schematic views illustrating the case
where a Mahalanobis distance is relatively small and the case where
the Mahalanobis distance is relatively large.
[0033] FIG. 17 is a schematic view of an apparatus shown in a
second exemplary embodiment of the present invention.
[0034] FIGS. 18A and 18B are views illustrating a spectrum image
and spectra, which have been used in the second exemplary
embodiment of the present invention.
[0035] FIGS. 19A, 19B, 19C and 19D are views illustrating an effect
of the selection of a feature value in the second exemplary
embodiment of the present invention.
[0036] FIG. 20 is an image illustrating a result of the
discriminant analysis which has been conducted in the second
exemplary embodiment of the present invention.
[0037] FIG. 21 is an image illustrating a result of the present
invention in the second exemplary embodiment of the present
invention.
[0038] FIG. 22 is a schematic diagram illustrating a concept of a
higher-order local autocorrelation (HLAC).
[0039] FIGS. 23A and 23B are images illustrating an effect of the
application of multidimensional information in the second exemplary
embodiment of the present invention.
DESCRIPTION OF THE EMBODIMENTS
[0040] Preferred embodiments of the present invention will now be
described in detail in accordance with the accompanying
drawings.
[0041] Embodiments of the present invention will be specifically
described below with reference to the flow charts and the drawings.
Incidentally, the following specific example is one example of
exemplary embodiments according to the present invention, but the
present invention is not limited to any such specific embodiment.
The present invention includes measuring a specimen having a
composition distribution in a space, and can be applied to results
provided by any measuring method as long as the measuring method
can obtain the information of a measured spectrum correlated with a
substance distributed within a biological tissue or a pathological
tissue contained in a lesion so that the information corresponds to
positional information at each point and positions of each point in
the space.
[0042] A view illustrated in FIG. 4 is a flow chart of image
reconstruction according to the present invention. The embodiment
will be described below with reference to the drawing according to
the order in this flow chart.
[0043] In the step of S101 in FIG. 4, a peak to be used in the
image reconstruction is selected. Here, the peak means a peak of
signal intensity in the case of the measured spectrum (for
instance, mass spectrum) as illustrated in FIG. 3A. On the other
hand, there is a spectroscopy which uses a spectrum in an
ultraviolet region, a visible region and an infrared region, or a
Raman spectroscopy which uses a Raman spectroscopic spectrum, as
the measured spectrum. The spectrum measured when such a
spectroscopy has been used forms a measured signal illustrated in
FIG. 3B. In this case, the signal intensity illustrated in FIG. 3C,
which has been provided by the discretization of the measured
signal, forms peaks of the signal intensity. Next, in the step of
S102, the data is normalized and digitized. In the step of S103,
multi-dimensional data is generated from the normalized and
digitized data, which is formed of positions of each point at which
the spectrum has been measured in the space and of a spectrum (peak
component) measured at each of the points in the space.
[0044] A view illustrated in FIG. 2 is a schematic view
illustrating the intensity distribution of the measured spectrum
which has been measured in each of the points on the space. For
instance, when a two-dimensional plane is considered as a space in
which signals are acquired, the information becomes
three-dimensional data. Each of the points of the three-dimensional
space from which these three-dimensional data are generated is
expressed by a coordinate (X, Y, Z). The components X and Y are
coordinates on the two-dimensional space (XY plane) in which the
measured spectrum signal is contained, and correspond to FIG. 2A.
The component Z is a measured spectrum signal at each of the points
on the XY plane, and corresponds to FIG. 2B. Accordingly, the
components X and Y store the X-coordinate and the Y-coordinate of
the point at which the signal has been measured, and the component
Z stores a value of the measured signal corresponding to the
intensity of each peak component.
[0045] In the step of S104 in FIG. 4, the signal is classified by
the generated classifier, and an image is output. Machine learning,
for instance, can be used for the generation of this classifier. In
this machine learning, a determination criterion is generated which
connects the measured data and the information on the biological
tissue, from already acquired data (which is referred to as
training data).
[0046] A view illustrated in FIG. 5 is a flow chart for generating
the classifier. The content will be described below with reference
to the drawing according to the order in this flow chart.
[0047] In the step of S201 in FIG. 5, a peak to be used in the
image reconstruction is selected. Next, in the step of S202, the
image data are divided into blocks. Here, the division into blocks
means that an image region is divided into each of a plurality of
small blocks. In the step of S203, the classifier is generated in
each block from the data of each of divided blocks, for instance,
by machine learning. As for a technique of the machine learning,
such methods can be used as a Fisher's linear discriminant method,
a SVM (Support Vector Machine), a decision tree, and a random
forest method which considers the ensemble average thereof. In the
step of S204, a classification model which can be applied to all of
the image regions is generated by a regression analysis of
classification conditions obtained in each of the image blocks.
Incidentally, this step may be omitted, and it is acceptable,
instead, to reconstruct the biological tissue image per every image
block by using the classifier generated in each image block, and
then integrate these images by interpolation processing and the
like to generate the biological tissue image in the image region.
The case will be described below where the Fisher's linear
discriminant method has been employed, as one example of supervised
machine learning. Incidentally, the discriminant analysis means the
Fisher's linear discriminant method, and the classification
conditions mean discriminant conditions which have been acquired by
the application of the Fisher's linear discriminant method.
[0048] The image region of an object may be all of the image
regions to be acquired, and may also be an image region which has
been partially selected. When the image region which has been
partially selected is the object, it is acceptable, for instance,
to previously set the image region which is not the object such as
an outer peripheral portion, in all of the acquired image regions,
and set the image region except for the previously set image
region.
[0049] FIGS. 6A to 6C illustrate a process of separating and
classifying a plurality of groups from spectrum data by the
discriminant analysis. A white frame in FIG. 6A shows a region in
which the spectrum data to be used as training data is acquired.
FIG. 6B is a schematic view of the spectrum data to be used. Each
spectrum of an object to be learned is accompanied by a
classification number (label) of the biological tissue, such as 1
for a cancer tissue and 0 for a normal tissue, for instance. FIG.
6C schematically illustrates such a state that a feature value
which has been acquired from the spectrum data is projected to a
feature space (classification space) and an optimal boundary line
is determined by the discriminant analysis. Here, the feature space
means a space to which the feature value is projected in order to
classify the attribute of the data, and the feature value means a
value suitable for classification, which is generated from original
data. A normalized peak intensity and the like can be considered as
the feature value in this case.
[0050] FIGS. 7A to 7C schematically illustrate a state of
inter-group dispersion and intra-group dispersion which are
projected components to a projection axis. FIG. 7B illustrates the
inter-group dispersion corresponding to a distance between gravity
centers of each group, and the inter-group dispersion is given by
Expression (1).
{w.sup.T( x.sub.1- x.sub.2)}.sup.2 [Expression 1]
[0051] In addition, FIG. 7C illustrates the intra-group dispersion
equivalent to dispersion within each group, and the intra-group
dispersion is given by Expression (2).
1 ? - 2 { ( ? - ? ) ? ? w + ( ? - 1 ) ? ? w } = ? ? S = 1 ? + ? - 2
{ ( ? - 1 ) ? + ( ? - 1 ) ? } ? indicates text missing or illegible
when filed [ Expression 2 ] ##EQU00001##
[0052] The vector w in the above Expression means a coefficient
vector shown in the following Expression (3). The vectors x.sub.1
and x.sub.2 in the above Expression mean a sample average vector of
each group shown by the following Expression (4). The matrices
S.sub.1 and S.sub.2 in the above Expression mean a sample-variance
covariance matrix of each group shown by the following Expression
(5). The expressions are expressions in the case where the feature
space is two dimensional, respectively. In addition, n.sub.1 and
n.sub.2 are the numbers of the data of each group.
w = ( w 1 w 2 ) [ Expression 3 ] ? = ( ? ) ? = ( ? ) [ Expression 4
] ? = ( ? ) S 2 = ( ? ) ? indicates text missing or illegible when
filed [ Expression 5 ] ##EQU00002##
[0053] The Fisher's linear discriminant method is a method of
determining an axis that maximizes a ratio of the inter-group
dispersion and the intra-group dispersion which are the projected
components to the axis, and such an axis is given by Expression
(6). In Expression (6), x represents a coordinate in a feature
space, and a position at which a reference numeral of H(x) changes
becomes a boundary that distinguishes both of the groups.
h ( x ) = ( ? - ? ) T S - 1 x - 1 2 ( ? - ? ) T S - 1 ( ? + ? ) ?
indicates text missing or illegible when filed [ Expression 6 ]
##EQU00003##
[0054] FIG. 7A schematically illustrates a classification axis
which is determined by the discriminant analysis.
[0055] FIGS. 8A to 8C schematically illustrate such a state that
confounding occurs when the discriminant analysis is conducted with
the use of data in a plurality of different image blocks. The
confounding means such a phenomenon that the data are mixed when
the data having different properties are used. The white frame in
FIG. 8A shows the block in the image of which the data is used.
FIG. 8B schematically illustrates spectrum data corresponding to
the blocks, and FIG. 8C schematically illustrates such a state that
the data are mixed due to confounding occurring when those data are
projected to the feature space.
[0056] FIGS. 9A to 9C schematically illustrate such a state that
the classification conditions (which are determined from Expression
(6)) for each image block are acquired by local management of the
data, the acquired classification conditions are subjected to the
regression analysis, and thereby classification models capable of
being applied to the image regions are acquired. Here, the local
management of the data means that the data is divided in such a
degree that confounding of the data does not occur. The white frame
in FIG. 9A shows the block in the image of which the data is used.
FIG. 9B illustrates such a state that the discriminant analysis is
applied to each of the image blocks. FIG. 9C illustrates such a
state that the regression analysis of the classification conditions
is conducted. Thus, the image is divided into an appropriate image
block size in such a degree that confounding does not occur. The
classification model is constructed which can be appropriately
applied to the image region, by the regression analysis of the
classification conditions that have been acquired from the
discriminant analysis of each of the blocks. Thereby, it is enabled
to conduct an appropriate classification while preventing the
confounding of the data. For information, the optimal image block
size can be determined, for instance, by using such a statistical
test as is given by Expression (7), a misclassification rate of the
training data, and the like.
? = ? - ? ? ? + ? ? ? indicates text missing or illegible when
filed [ Expression 7 ] ##EQU00004##
[0057] Here, .sigma..sub.1 and .sigma..sub.2 in Expression (7) mean
a sample variance of each group. In addition, z.sub.0 is a test
value, and the block size is determined so that the value becomes a
constant value or more (for instance, 1.96 or more).
[0058] FIGS. 10A to 10C schematically illustrate a series of
processes illustrated in the flow charts in FIG. 4 and FIG. 5. In
FIG. 10A, the classification model is generated by the machine
learning and the regression analysis, and in FIG. 10B, data which
have been newly measured are input. Then, in FIG. 10C, a
distribution image (which is obtained from result of machine
learning) of the biological tissue distribution, for instance, is
acquired as a reconstruction image.
[0059] In addition, the data to be used in the machine learning and
the classification may not only be spectrum data of each point in
the space, but also both the spectrum data of each point in the
space and the distribution information (morphological information)
of each spectrum component, for instance, may be used.
[0060] In this case, a peripheral area of a pixel which receives
attention, for instance, is cut out, and attention is paid to a
pattern which the region forms. For instance, when the
two-dimensional plane is considered as a space of which the signal
is acquired, the data to be used in the machine learning and the
classification shall be data having a three-dimensional structure
in a total of the distribution information and the spectrum
information in the plane (which is referred to as multi-dimensional
information).
[0061] The procedure of the machine learning and the classification
in the case where the multi-dimensional information has been used
is essentially the same as that in the case where the above
described spectrum data has been used. However, in this case, the
data itself is not used for a vector (which is referred to as
feature vector hereafter) for use in the classification, but also
it is possible to acquire a plurality of feature values suitable
for describing the pattern, define the feature values as a feature
vector, and use the feature vector for the machine learning and the
classification processing. As a representative example of the
feature value, there are a volume, a curvature, a space gradient,
HLAC (high-order local autocorrelation) and the like. Here, the
high-order autocorrelation function of N-order is defined by
Expression (8) for displacement directions (a.sub.1, a.sub.2, . . .
, a.sub.N), when the image of an object is represented by f(r).
x.sub.N(a.sub.1, a.sub.2, . . . , a.sub.N)=.intg.f(r)f(r+a.sub.1) .
. . f(r+a.sub.N)dr
[0062] In addition, the high-order local autocorrelation function
is defined so that the displacement directions are limited to a
localized area of a reference point r (for instance, 3.times.3
pixels around reference point r). FIG. 22 illustrates a reference
pattern in the case of 0-order and 1-order. A pixel with a charcoal
gray becomes a center point for a reference when the
autocorrelation is calculated.
[0063] In addition, it is also possible to select the feature value
to be used in the machine learning beforehand. In this case, for
instance, it is acceptable to calculate a Mahalanobis distance
which is obtained by projecting each of the feature values to the
feature space and is defined by the ratio of the inter-group
dispersion and the intra-group dispersion of each group, and to
select the feature value for use in the classification. The result
that the Mahalanobis distance is small corresponds to the case as
in FIG. 16A, and the result that the Mahalanobis distance is large
corresponds to the case as in FIG. 16B, when illustrated by
Comparative Examples. If the Mahalanobis distance is large, the
classification becomes easier. Accordingly, it is also possible to
preferentially select such a feature value that the Mahalanobis
distance between each group that receives attention is large.
[0064] The present invention can be achieved by an apparatus which
carries out the above described specific embodiment. FIG. 1
illustrates one example of the configuration of the whole apparatus
on which the present invention is mounted. A specimen 1 on a
substrate and a detector 2 for a signal are shown. In addition, a
signal processing apparatus 3 which conducts the above described
processing for the acquired signal, and an image display apparatus
4 which displays the signal processing result on a screen are
shown.
[0065] The configuration will be more specifically described while
taking a TOF-SIMS as an example. In the configuration, the detector
2 measures secondary ions (which are shown by dotted line in FIG.
1) which are generated in the specimen 1 that has been irradiated
with primary ions (not-shown), and transmits the signal which has
been converted into an electrical signal, to the signal processing
apparatus 3. For information, the type of the primary ion is not
limited, and a usable detector includes not only a detector for one
dimension but also a semiconductor detector for two dimension.
Furthermore, it is possible to use a laser in place of the primary
ion, and also to use a specimen stage having a scanning function
together. The measured data has a three-dimensional data structure
in which a mass spectrum is stored in a coordinate point on the XY
plane of the specimen 1. In addition, when the data has been
integrated, the measured data becomes four-dimensional data.
However, the integrated data becomes three dimensional, and can be
subjected to similar processing.
[0066] In addition, FIG. 17 also illustrates one example of the
configuration of the apparatus on which the present invention is
mounted. A light source 11 and an optical system 12 are shown. In
addition, the specimen 1 to be measured, a stage 14 on which the
specimen is arranged, and the detector 2 for a signal are shown. In
addition, the signal processing apparatus 3 which subjects the
acquired signal to the above described processing, and the image
display apparatus 4 which displays the signal processing result on
the screen are shown.
[0067] In FIG. 17, a measurement system of a transmission type of
arrangement is shown, but a reflection type of arrangement is also
possible. In addition, ultraviolet rays, visible light, infrared
rays and the like can be used as a light source. The detector also
includes a single detector, a line-shaped detector and a
two-dimensional detector, and the type is not limited. Furthermore,
such a method is also acceptable as to combine an interferometer
with the apparatus and acquire a spectrum through Fourier
transformation or Laplace transformation. It is also possible to
add the scanning function to the specimen stage. The measured data
has a three-dimensional data structure in which the spectrum is
stored in the coordinate point on the XY plane of the specimen 1.
In addition, when the data has been integrated, the measured data
becomes four-dimensional data. However, the integrated data becomes
three dimensional, and can be subjected to similar processing.
[0068] In addition, FIG. 17 includes also nonlinear spectroscopy
such as coherent anti-Stokes Raman scattering (CARS, Coherent
Anti-Stokes Raman Scattering) and stimulated Raman scattering (SRS,
Stimulated Raman Scattering).
[0069] Furthermore, the signal processing apparatus and the image
output apparatus (that handle signal after detector) which are
features of the present invention can be also applied to the
configuration, as long as the apparatus has such a structure that
the spectrum is stored in the coordinate point on one particular
cross section (which has constant thickness) of the specimen.
Specifically, the apparatuses can be applied also to a
two-dimensional spectrum measuring system with the use of X-rays, a
terahertz wave, an electromagnetic wave or the like.
Exemplary Embodiment 1
[0070] Exemplary Embodiment 1 of the present invention will be
described below. In the present exemplary embodiment, a tissue
slice (product made by Pantomics, Inc.) of an expression level 2+
of a HER2 protein which had been subjected to trypsin digestion
processing was subjected to an SIMS measurement on the following
conditions, in which a TOF-SIMS 5 type apparatus (trade name) made
by ION-TOF GmbH was used.
[0071] Primary ion: 25 kV Bi.sup.+, 0.6 pA (pulse current value),
macro-raster scan mode
[0072] Pulse frequency of primary ion: 5 kHz (200 .mu.s/shot)
[0073] Pulse width of primary ion: approximately 0.8 ns
[0074] Beam diameter of primary ion: approximately 0.8 .mu.m
[0075] Measurement range: 4 mm.times.4 mm
[0076] Pixel number in measurement of secondary ion:
256.times.256
[0077] Integration period of time: 512 shots for one pixel, one
time scan (approximately 150 minutes)
[0078] Detection mode for secondary ion: positive ion
[0079] The XY coordinate information which shows the positions for
each measurement pixel and the mass spectrum in one shot are
recorded in the obtained SIMS data. For instance, each of the
measurement pixels contains the information on the peak
(m/z=720.35) which corresponds to a mass number of a molecule in
one of digestive fragments of the HER2 protein, to which one sodium
atom adsorbs, and on the peak components originating in each
biological tissue, as the spectrum data.
[0080] FIG. 11A illustrates an image obtained through the
observation of the tissue slice (product made by Pantomics, Inc.)
of the expression level 2+ of the HER2 protein, of which the HER2
protein was subjected to immunostaining, by an optical microscope.
In FIG. 11A, the portion in which there are more expressions in the
HER2 protein is indicated whiter. In addition, the specimen that
was used in the SIMS measurement and the specimen that was
subjected to the immunostaining are adjacent slices to each other,
which were cut out from the same diseased tissue (paraffin block),
and are not identical.
[0081] FIG. 11B illustrates a spectrum which was measured in an
image block surrounded by the white frame in FIG. 11A. FIG. 11C
illustrates a result of the discriminant analysis that was
conducted for two peak components (values of corresponding m/z are
692.35 and 1101.5, respectively), which were selected from the
spectrum of FIG. 11B. It is understood in FIG. 11C that the
different groups can be clearly separated from each other.
[0082] The white frames in FIG. 12A illustrate a plurality of image
blocks. FIG. 12B illustrates a result of the discriminant analysis
which was conducted for the spectrum data of each of the image
blocks. It is understood in FIG. 12B that the characteristics of
the data change according to the positions of the image blocks, and
the classification condition also changes according to the change
of the characteristics.
[0083] The white frame in FIG. 13A illustrates an image block
formed by a combination of the plurality of the image blocks in
FIG. 12A. FIG. 13B illustrates a result of the plotting of the data
in the white frame in FIG. 13A on the feature space. It is
understood that such a phenomenon that data in different groups are
mixed with each other, which is so-called confounding, occurs in
the white frame in FIG. 13B.
[0084] FIG. 14A illustrates a result of the division of the image
region into image blocks. FIG. 14B illustrates a result of the
regression analysis which was conducted for the classification
conditions based on the result of the discriminant analysis that
was conducted for a plurality of the image blocks. From the result
of this regression analysis, a classification model is generated
which can be applied to the image region.
[0085] FIG. 15A illustrates a result of the reconstruction of the
image, by the application of the classification conditions obtained
from the discriminant analysis of a single image block, to the
image region. In addition, FIG. 15B illustrates a result of the
reconstruction of the image, by the application of the discriminant
analysis of the divided blocks according to the present invention,
to the image region. As is understood when the inside of the white
frame is referred and compared to FIG. 11A of reference, a
biological tissue image with higher precision can be acquired
according to the present invention.
Exemplary Embodiment 2
[0086] Exemplary Embodiment 2 of the present invention will be
described below. In the following exemplary embodiment, a mouse
liver tissue was measured with the use of a microscope which uses
stimulated Raman scattering. The power of a TiS laser used as a
light source was 111 mW, and the intensity of an Yb fiber laser was
127 mW before the laser was incident on an object lens. The mouse
liver tissue of the specimen was subjected to formalin fixation
treatment, and was cut into a thin slice with a thickness of 100
micrometers. The tissue slice was subjected to measurement in a
state of being embedded in a glass together with a PBS buffer. The
measurement range is a 160 micrometers square, and the data
measured 10 times were integrated. The image data was a 500 pixels
square, and the measurement period of time was 30 seconds.
[0087] The XY coordinate information which shows the positions of
each measurement pixel, and the spectrum information in each
coordinate are recorded in the obtained spectral image data. For
instance, each of the measurement pixels contains the information
on the peak components originating in the components of the tissue
constituting the specimen, as the spectrum data. In addition, the
measurement of the spectrum data was conducted with a sampling
interval of 1 kayser (1 cm.sup.-1).
[0088] FIG. 18A is an image formed by adding up of signals in all
measured spectral regions, which are based on a measurement result
of a liver tissue. FIG. 18B is a graph obtained from picked-up
spectra of portions corresponding to the cell nucleus, the cell
cytoplasm and the erythrocyte; and in the graph, a horizontal axis
corresponds to a wave number (while the numerical value in the
graph is an index for distinguishing the wave number and the index
will be referred to hereafter), and a vertical axis corresponds to
signal intensity. It is understood that spectrum signals different
among each tissue are obtained, as is illustrated in FIG. 18B.
[0089] FIG. 19A is a graph obtained from the calculation of a
Mahalanobis distance between the cell nucleus (group 1) and the
cell cytoplasm (group 2) for each wave number. It is understood
that the Mahalanobis distance is large when the indices are values
between 7 and 8. FIG. 19B is a graph obtained by determining the
spectrum intensities corresponding to the indices 7 and 8 as the
feature values, and plotting one part of the training data onto the
two-dimensional feature space. It is understood that the group 1
and the group 2 can be clearly distinguished. FIG. 19C is a graph
obtained from the calculation of a Mahalanobis distance between the
cell cytoplasm (group 2) and the erythrocyte (group 3) for each
wave number. It is understood that the Mahalanobis distance is
large when the indices are values between 15 and 17. FIG. 19D is a
graph obtained similarly by determining the spectrum intensities
corresponding to the indices 15 and 16 as the feature values, and
plotting one part of the training data onto the two-dimensional
feature space. It is understood that the group 2 and the group 3
can be clearly distinguished. On the other hand, it is understood
that the group 1 and the group 2 result in resisting being
distinguished. In such a case, it is acceptable to use all of the
feature values which are suitable for classification among each of
the groups, and then project the resultant feature values to the
feature space. In this case, it is acceptable, for instance, to
determine the spectrum intensities corresponding to the indices 7,
8 and so on, and 15, 16 and so on as the feature values, project
the resultant feature values to the multidimensional feature space,
and classify each group.
[0090] FIG. 20 is a result of the discriminant analysis which has
been conducted for the plotting of spectrum intensities
corresponding to the index 8 and the index 15 on a two-dimensional
space, with the use of training data that correspond to the cell
nucleus (group 1), the cell cytoplasm (group 2) and the erythrocyte
(group 3), respectively. It is understood that each group can be
clearly separated from each other.
[0091] FIG. 21 is an image obtained by classifying the cell
nucleus, the cell cytoplasm and the erythrocyte based on the result
of the previously described discriminant analysis, and
reconstructing the resultant image. It is understood that each
tissue is appropriately classified and is color-coded. Thus, the
present invention can classify a structure in a biological tissue
without dyeing.
[0092] FIG. 23A illustrates a result of classification processing
that has been conducted with the use of only spectrum intensity as
a feature value; and FIG. 23B illustrates a result of
classification processing that has been conducted with the use of
the spectrum intensity and HLAC in 0-order and 1-order, which is
calculated from the distribution of the spectrum intensity, as a
feature value. It is understood that an outline of a structure in
each tissue is more clearly drawn in FIG. 23B compared to that in
FIG. 23A. Thus, the outline of the structure in a biological tissue
can be more clearly drawn by the utilization of such
multidimensional information.
[0093] The present invention can be used as a tool which more
effectively supports a pathological diagnosis.
[0094] The method according to the present invention can
reconstruct a biological tissue image, by measuring a spatial
distribution of a measured spectrum, using both measured spectrum
information thereof and morphological information which is obtained
from the distribution information of peak components and applying
machine learning to the information. Furthermore, even in the case
where characteristics of data change and classification conditions
change due to a difference among measured positions of the measured
spectrum and the like, at this time, the method can reconstruct a
biological tissue image by employing appropriate classification
conditions. Thereby, the biological tissue can be classified with
higher precision compared to a conventional method, and accordingly
the method is useful when being applied to the pathological
diagnosis or the like.
[0095] While the present invention has been described with
reference to exemplary embodiments, it is to be understood that the
invention is not limited to the disclosed exemplary embodiments.
The scope of the following claims is to be accorded the broadest
interpretation so as to encompass all such modifications and
equivalent structures and functions.
[0096] This application claims the benefit of Japanese Patent
Applications No. 2013-000883, filed Jan. 8, 2013, No. 2013-163399,
filed Aug. 6, 2013 and No. 2013-251050, filed Dec. 4, 2013 which
are hereby incorporated by reference herein in their entirety.
* * * * *