JP2014134517A - Pathologic tissue image analysis method, pathologic tissue image analyzer and pathologic tissue image analysis program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To present quantitation information so as to be easily utilized in clinical diagnosis.SOLUTION: A sinusoidal area extraction processing part 210 extracts a sinusoidal area in an analysis object area from a pathologic tissue image acquired by an image acquisition processing part 10, an interstitial area extraction processing part 220 extracts an interstitial area in the analysis object area, a cell cord extraction processing part 230 extracts an area other than the extracted sinusoidal area and interstitial area as a cell cord area in the analysis object area from the pathologic tissue image, and a quantitation processing part 240 quantifies a pathologic feature amount in each cell cord.

Description

  The present invention relates to a pathological tissue image analysis method, a pathological tissue image analysis apparatus, and a pathological tissue image analysis program that support pathological tissue image diagnosis.

  In the medical field, as part of pathological diagnosis, a part of an organ is collected from a patient as a specimen, and this is sliced into several micron-thick tissue slice glass slides (preparations) and observed under a microscope. Biopsy tissue diagnosis is known in which diagnosis is performed based on morphological information such as changes in the size and shape of the nucleus and changes in pattern as a tissue, and staining information. In this case, since the sliced specimen hardly absorbs and scatters light and is almost colorless and transparent, it is general that the specimen is stained with a dye prior to observation.

  It is necessary for a doctor to make a pathological diagnosis for diagnosing a lesion site by observing a pathological slide with a microscope and taking macroscopic findings into consideration, and the amount of work for that is enormous, and the burden on the doctor is heavy. Therefore, in order to efficiently perform the observation work for diagnosis and reduce the burden on the doctor, development of various computer diagnosis support devices (hereinafter referred to as CAD: Computer-Aided Diagnosis) using image processing technology by a computer system. Has been done.

  For example, the liver pathological image shown in FIG. 35 is utilized as a golden standard in various scenes such as rapid diagnosis during surgery, estimation of the progression of liver fibrosis, hepatocellular carcinoma from tissue obtained by biopsy, and pathological anatomy. . In addition, in recent years, digitization of pathological images has progressed, and a hole slide image (WSI: Whole Slide Image) in which a specimen can be observed while virtually changing the magnification freely is becoming widespread. WSI, which uses a digital microscope to photograph whole pathological glass specimens, is becoming popular in various situations such as remote diagnosis, education, and presentation. On the other hand, expectation for CAD using digitized pathological images is increasing.

  Conventionally, as a conventional method for supporting computer diagnosis of pathological images stained with, for example, hematoxylin and eosin (HE), for example, a system for detecting a region of interest such as cancer (see, for example, Patent Document 1 and Patent Document 2), a cell nucleus region, etc. In addition, a system (for example, refer to Patent Document 3 and Patent Document 4) that measures basic feature amounts indicating features of respective shapes of a cytoplasmic region and a glandular cavity region, extracts a target region, and presents information has been proposed.

  For example, Patent Document 5 proposes a system that extracts a gland duct region from a low-magnification pathological image, calculates a feature amount related to the gland duct structure from the gland duct region, and diagnoses and determines a disease state. In the disclosed technique of Patent Document 5, when extracting the gland duct region and calculating the distance of the gland duct from the distance conversion image, the gland duct is thinned, and the distance from the center line to the boundary is regarded as the radius and doubled. By doing so, the thickness is calculated.

In Patent Document 6, an optimal value is determined by a neural network from a plurality of candidate values obtained by combining a plurality of automatic threshold selection methods, and the threshold is set. There has been proposed a system for analyzing an image after binarizing the image. In the technique disclosed in Patent Document 6, an optimal threshold value for binarization is calculated in a MIB-1 image by a plurality of methods, cell nuclei are extracted, and a labeling index is automatically calculated.
Further, in Non-Patent Document 1, feature quantities mainly centering on cell nuclei are calculated from liver pathological specimen images to estimate Edmonson grade.

  Further, as a method for extracting sinusoids from a hepatocyte pathological image, Non-Patent Document 2 discloses a method of connecting and extracting partial discontinuities of sinusoids by selectively convolving a filter in accordance with a luminance gradient. .

JP 2011-215061 A JP 2012-179336 A JP 2009-180539 A JP 2009-115598 A JP 2009-229203 A JP 2004-199391 A

P. Huang and Y. Lai, “Effective segmentation and classification for hcc biopsy images,” Pattern Recogni-tion 43 (4), pp. 1550-1563, 2010. M. Ishikawa, AS Taha, F. Kimura, M. Yamaguchi, H. Nagahashi, A. Hashiguchi, M. Sakamoto, “An Orientation Selective Filter Based Approach to Extract Sinusoids from HE-Stained Liver Specimens,” IEICE technical report, 111 (389), 77-157 (2012) (Japanese). Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Susstrunk, SLIC Superpixels Compared to State-of-the-art Superpixel Methods, accepted to the IEEE Transactions on Pattern Analysis and Machine Intelligence, May 2012. F. Tomoya, Y. Masahiro, M. Yuri, K. Fumikazu, H. Akinori, S. Michiie, and O. Nagaaki, “Calculation quantitative feature amount in digital slide image of pathological specimen,” IEICE technical report Japanease, 2011. Tomoharu Kiyuna, Akira Saito, Atsushi Marugame, Yoshiko Yamashita, Maki Ogura, NEC Corp. (Japan); Eric Cosatto, NEC Labs. America, Inc. (United States); Tokiya Abe, Akinori Hashiguchi, Michiie Sakamoto, Keio Univ. Japan), “Automatic classification of hepatocellular carcinoma images based on nuclear and structural features”, SPIE Medical Imaging 2013, 2.2013.

  However, the main purposes of the disclosed techniques of Patent Documents 1 to 4 are all to call attention by extracting a region of interest and presenting it to a doctor. The disclosed technique of Patent Document 1 uses a gland duct as a target, and the target pathological images are stomach cancer, breast cancer, and prostate cancer, and not liver cancer. Basically, the ducts are extracted in a circular shape and can be evaluated individually. The purpose of the cell cord is different from this method, which has continuity and quantifies it by the label unit of the thinning result.

  The counting of pathological images is an issue already addressed by special staining. For example, in the special staining ki67, the positive nuclei are stained brown and the negative nuclei are stained blue, so the ratio of positive nuclei is actually counted to calculate the labeling index. For example, as disclosed in Patent Document 6, the Labeling index has a high affinity with computer diagnosis support because the number of positive nuclei is dyed brown by special staining and research is progressing.

  However, liver pathology images are not diagnosed with a single index like the labeling index, but are diagnosed taking into account various structures such as the NC ratio (the ratio of cell nucleus to cytoplasm), the size of the cell nucleus, and the structure of the cell cord. The

  In the disclosed technique such as Patent Document 1 described above, pathological images are divided and evaluated within a certain range, but the structure is sufficiently considered for pathological images of livers having a cord-like structure including sinusoids and stroma. I can't say that.

  Accordingly, an object of the present invention is to provide a pathological tissue image analysis method, a pathological tissue image analysis apparatus, and a pathological tissue image analysis program capable of presenting quantification information so that it can be easily used in clinical diagnosis in view of the conventional situation as described above. And providing diagnostic support for pathological images.

  In particular, by presenting the quantified information that enables quantitative pathological diagnosis by analyzing and quantifying complex structures in HE-stained pathological images of the liver so that they can be used in clinical diagnosis. An object of the present invention is to provide a pathological tissue image analysis method, a pathological tissue image analysis apparatus, and a pathological tissue image analysis program that can perform diagnosis support for pathological images.

  Other objects of the present invention and specific advantages obtained by the present invention will become more apparent from the description of embodiments described below.

  In the present invention, the region other than the blank region and the interstitial region is extracted from the pathological tissue image obtained by imaging the analysis target by extracting the region where the cells do not exist and the interstitial region in the analysis target region. Extracted as a parenchymal cell region in the analysis target region, and the pathological feature amount is quantified for each aggregate of the parenchymal cells extracted as the parenchymal cell region.

  That is, the present invention is a pathological tissue image analysis method, an image capture processing step for capturing a pathological tissue image obtained by imaging an analysis target, and cells in an analysis target region from the captured pathological tissue image. A blank area extraction process step for extracting a non-existing area; a stroma area extraction process step for extracting a stromal area in the analysis target area from the pathological tissue image; an area other than the extracted blank area and the interstitial area A parenchymal cell region extraction processing step for extracting a pathological feature amount from the pathological tissue image as a parenchymal cell region in the analysis target region, and a quantification process for quantifying a pathological feature value for each aggregate of parenchymal cells extracted as the parenchymal cell region And a step.

  The pathological tissue image analysis method according to the present invention may further include, for example, a display step of displaying the pathological feature value quantified in the quantification processing step.

  In the pathological tissue image analysis method according to the present invention, in the display step, for example, an index or a label representing a lesion state is displayed using a pathological feature value digitized for each aggregate of the parenchymal cells. can do.

  The pathological tissue image analysis method according to the present invention further includes a region-of-interest mask generation step for generating mask information representing a region to be noted in the pathological tissue image, and in the display step, The statistical value of the feature value in the attention area mask can be displayed using the pathological feature value digitized for each aggregate.

  Further, in the pathological tissue image analysis method according to the present invention, in the quantification processing step, the pathological feature amount is quantified as a group of the parenchymal cells in units of a series of cells having a predetermined length or less. Can do.

  Further, in the histopathological image analysis method according to the present invention, in the quantification processing step, the pathological feature amount is quantified based on a unit with a branch of a series of cells as a boundary as the aggregate of the parenchymal cells. can do.

  Further, the histopathological image analysis method according to the present invention includes a storage step of storing the pathological feature value quantified for each assembly of the parenchymal cells together with the label information assigned to each assembly. Can do.

  In the pathological tissue image analysis method according to the present invention, the pathological tissue image may be an image of a hepatocyte subjected to a staining process.

  Further, the pathological tissue image analysis method according to the present invention includes an sinusoidal region extraction processing step of extracting an sinusoidal region in the analysis target region from the captured pathological tissue image as the blank region extraction processing step, In the cell region extraction processing step, the extracted region other than the sinusoidal region and the stroma region is extracted from the pathological tissue image as a substantial cell region in the analysis target region, and in the quantification processing step, the parenchymal cell is extracted. The pathological feature amount can be quantified for each cell line extracted as a region.

  In the pathological tissue image analysis method according to the present invention, the sinusoidal region extraction processing step includes a boundary enhancement processing step of performing boundary enhancement processing using a direction selection filter for the pathological tissue image of the analysis target region, and the boundary enhancement processing Clustering processing step for extracting the sinusoidal region by clustering the pathological tissue image of the analysis target region that has been subjected to the analysis, and reduction of false extraction regions by the support vector machine (SVM) for the extracted sinusoidal region And an erroneous extraction area reduction processing step for performing processing.

  In the pathological tissue image analysis method according to the present invention, the interstitial region extraction processing step includes a fiber probability calculation processing step for performing fiber probability calculation processing based on texture characteristics for the pathological tissue image of the analysis target region, and the analysis target region. Superpixelization processing step for superpixel processing to cluster the similar pathological tissue images in similar colors, and calculating the average fiber probability for each superpixel and extracting superpixels with high fiber probability as stroma A first stromal extraction processing step and a second stromal extraction processing step for extracting superpixels with a large number of lymphocytes as stroma.

  Further, in the pathological tissue image analysis method according to the present invention, in the quantification processing step, the center of the closed region graphic of the parenchymal cell region is obtained for the pathological tissue image of the parenchymal cell region extracted in the parenchymal cell region extraction processing step. It is possible to perform a distance measurement process in which the width along the line is measured and output as a quantified pathological feature amount.

  In the pathological tissue image analysis method according to the present invention, the quantification processing step includes a thin line by obtaining a center line of the parenchymal cell region with respect to the pathological tissue image of the parenchymal cell region extracted in the parenchymal cell region extraction processing step. A thinning process step, a cell line classification processing step for classifying the parenchymal cell area in branches, with branches from the branching of the thinned parenchymal cell area as branches, and a parenchyma of the classified branch unit The pathological tissue image of the cell region may include a distance measurement processing step of measuring a distance from the center line to the boundary of the parenchymal cell region and outputting as a quantified pathological feature amount.

  Furthermore, in the pathological tissue image analysis method according to the present invention, the quantification processing step may further calculate and output the nucleus density for each cell cord or the average of the NC ratio for each cord as the pathological feature amount. it can.

  The present invention is a histopathological image analysis apparatus, an image capture processing unit that captures a pathological tissue image obtained by imaging an analysis target, and an analysis from the pathological tissue image captured by the image capture processing unit Extracted by a blank area extraction processing unit that extracts a region in which no cells exist in the target region, a stromal region extraction processing unit that extracts a stromal region in the analysis target region from the pathological tissue image, and the stromal region extraction processing unit A parenchymal cell region extraction processing unit that extracts a region other than the blank region and the stromal region extracted by the stromal region extraction processing unit from the pathological tissue image as a parenchymal cell region in the analysis target region, and the parenchymal cell And a quantification processing unit that quantifies a pathological feature amount for each aggregate of parenchymal cells extracted as a parenchymal cell region by the region extraction processing unit.

  The pathological tissue image analysis apparatus according to the present invention may further include a display unit that displays the pathological feature value quantified by the quantification processing unit.

  Further, in the histopathological image analysis apparatus according to the present invention, the quantification processing unit displays the index or the label representing the state of the lesion using the pathological feature value quantified for each aggregate of the parenchymal cells. It can be displayed by the part.

  Further, in the pathological tissue image analysis apparatus according to the present invention, the quantification processing unit generates mask information representing a region to be noted in the pathological tissue image, and is quantified for each aggregate of the parenchymal cells. By using the pathological feature amount, the statistical value of the feature amount in the attention area mask can be displayed on the display unit.

  Further, in the histopathological image analysis apparatus according to the present invention, the quantification processing unit quantifies a pathological feature amount as a group of the parenchymal cells in units of a series of cells having a predetermined length or less. Can do.

  Further, in the histopathological image analysis apparatus according to the present invention, the quantification processing unit quantifies a pathological feature amount as a group of the parenchymal cells based on a unit having a cell branch as a boundary. can do.

  The pathological tissue image analysis apparatus according to the present invention further includes an information storage unit that stores the pathological feature value quantified for each assembly of parenchymal cells together with label information assigned to each assembly. It can be.

  In the pathological tissue image analysis apparatus according to the present invention, the pathological tissue image may be an image of a hepatocyte subjected to a staining process.

  Further, in the pathological tissue image analysis apparatus according to the present invention, the blank region extraction processing unit is an sinusoidal region extraction processing unit that extracts an sinusoidal region in the analysis target region from the captured pathological tissue image, and the parenchymal cell In the region extraction processing unit, the region other than the sinusoidal region and the stroma region extracted by the sinusoidal region extraction processing unit is extracted from the pathological tissue image as a substantial cell region in the analysis target region, and the quantification is performed. The processing unit can quantify the pathological feature amount for each cell line extracted as a substantial cell region by the substantial cell region extraction processing unit.

  Further, in the pathological tissue image analysis apparatus according to the present invention, the sinusoidal region extraction processing unit includes a boundary enhancement processing unit that performs boundary enhancement processing using a direction selection filter on the pathological tissue image of the analysis target region, and the boundary enhancement processing A clustering processing unit that performs a sinusoidal region extraction process by clustering on the pathological tissue image of the analysis target region that has undergone boundary enhancement processing by a unit, and a support vector machine (SVM) for the sinusoidal region extracted by the clustering processing unit : Support Vector Machine) may include an erroneous extraction area reduction processing step for performing erroneous extraction area reduction processing.

  In the pathological tissue image analysis apparatus according to the present invention, the stroma region extraction processing unit includes a fiber probability calculation processing unit that performs a fiber probability calculation process based on texture characteristics on the pathological tissue image of the analysis target region, and the analysis. Superpixelization processing unit that performs superpixel processing to cluster pathological tissue images of the target area into regions with similar colors, and interstitial superpixels with high fiber probability by calculating the average fiber probability for each superpixel And a second interstitial extraction processing unit that extracts superpixels with many lymphocytes as stroma.

  Further, in the pathological tissue image analysis apparatus according to the present invention, the quantification processing unit includes a center of a closed region graphic of the parenchymal cell region with respect to the pathological tissue image of the parenchymal cell region extracted by the parenchymal cell region extraction processing unit. It is possible to perform a distance measurement process in which the width along the line is measured and output as a quantified pathological feature amount.

  In the pathological tissue image analysis apparatus according to the present invention, the quantification processing unit obtains a center line of the parenchymal cell region for the pathological tissue image of the parenchymal cell region extracted by the parenchymal cell region extraction processing unit. A thinning processing unit that thins the cell line, a cell line classification processing unit that classifies the parenchymal cell region in branches by branching from the branch to the branch of the thinned parenchymal cell region, and the cell cord classification processing unit. A distance measurement processing unit that measures the distance from the center line of the parenchymal cell region to the boundary and outputs the quantified pathological feature amount for the classified pathological tissue image of the parenchymal parenchymal cell region; can do.

  Furthermore, in the pathological tissue image analysis apparatus according to the present invention, the quantification processing unit may further calculate and output the nuclear density for each cell cord or the average of the NC ratio for each cord as the pathological feature amount. it can.

  The present invention relates to a pathological tissue image analysis program executed by a computer included in a pathological tissue image analysis apparatus that analyzes a pathological tissue image obtained by imaging an analysis target, and from the captured pathological tissue image A blank area extraction processing unit that extracts a region in which no cells exist in the analysis target region, a stroma extraction region extraction processing unit that extracts a stroma region in the analysis target region from the pathological tissue image, and the stroma region extraction processing unit A parenchymal cell region extraction processing unit that extracts from the pathological tissue image a region other than the blank region extracted by the stromal region extraction processing unit and a region other than the interstitial region extracted from the pathological tissue image in the analysis target region, and As a quantification processing unit for quantifying the pathological feature amount for each assembly of parenchymal cells extracted as a parenchymal cell region by the parenchymal cell region extraction processing unit Characterized in that the functioning of the computer.

  The pathological tissue image analysis program according to the present invention may further display the pathological feature value quantified by the quantification processing unit on the display unit.

  Further, in the histopathological image analysis program according to the present invention, the quantification processing unit displays the index or the label representing the state of the lesion using the pathological feature value quantified for each aggregate of the parenchymal cells. It can be displayed by the part.

  Further, in the pathological tissue image analysis program according to the present invention, the quantification processing unit generates mask information representing a region to be noted in the pathological tissue image, and is quantified for each aggregate of the parenchymal cells. By using the pathological feature amount, the statistical value of the feature amount in the attention area mask can be displayed on the display unit.

  Further, in the histopathological image analysis program according to the present invention, the quantification processing unit quantifies a pathological feature amount as a group of the parenchymal cells in units of a series of cells having a predetermined length or less. Can do.

  Further, in the histopathological image analysis program according to the present invention, the quantification processing unit quantifies the pathological feature amount as a group of the parenchymal cells based on a unit having a boundary of a series of cells as a boundary. can do.

  The pathological tissue image analysis program according to the present invention further stores the pathological feature value quantified for each assembly of the parenchymal cells in an information storage unit that stores the pathological feature amount together with the label information assigned to each assembly. Can be.

  In the pathological tissue image analysis program according to the present invention, the pathological tissue image may be an image of a hepatocyte subjected to a staining process.

  In the pathological tissue image analysis program according to the present invention, the blank region extraction processing unit is an sinusoidal region extraction processing unit that extracts an sinusoidal region in the analysis target region from the captured pathological tissue image, and the parenchymal cell In the region extraction processing unit, the region other than the sinusoidal region and the stroma region extracted by the sinusoidal region extraction processing unit is extracted as a substantial cell region in the analysis target region from the pathological tissue image, and is quantified. The processing unit can quantify the pathological feature amount for each cell line extracted as a substantial cell region by the substantial cell region extraction processing unit.

  In the pathological tissue image analysis program according to the present invention, the sinusoidal region extraction processing unit includes a boundary enhancement processing unit that performs a boundary enhancement process on a pathological tissue image of the analysis target region using an orientation selection filter, and the boundary enhancement process. A clustering processing unit that performs a sinusoidal region extraction process by clustering on the pathological tissue image of the analysis target region that has undergone boundary enhancement processing by a unit, and a support vector machine (SVM) for the sinusoidal region extracted by the clustering processing unit The computer can be caused to function as a sinusoidal region extraction processing unit including an erroneous extraction region reduction processing step for performing erroneous extraction region reduction processing by (Support Vector Machine).

  Further, the pathological tissue image analysis program according to the present invention provides a fiber probability calculation processing unit that performs a fiber probability calculation process based on texture features for the pathological tissue image of the analysis target region, and a color for the pathological tissue image of the analysis target region. A superpixel processing unit that performs superpixel processing for clustering into regions similar to each other, and a first interstitial extraction that calculates an average fiber probability for each superpixel and extracts a superpixel having a high fiber probability as a stroma The computer can be caused to function as a stromal region extraction processing unit that includes a processing unit and a second stromal extraction processing unit that extracts superpixels with many lymphocytes as a stroma.

  Further, in the pathological tissue image analysis program according to the present invention, the quantification processing unit includes the center of the closed region figure of the parenchymal cell region with respect to the pathological tissue image of the parenchymal cell region extracted by the parenchymal cell region extraction processing unit. It is possible to perform a distance measurement process in which the width along the line is measured and output as a quantified pathological feature amount.

  The pathological tissue image analysis according to the present invention is a thinning processing unit that thins the pathological tissue image of the parenchymal cell region extracted by the parenchymal cell region extraction processing unit by obtaining a center line of the parenchymal cell region. A branch from the thinned parenchymal cell region to a branch, and a cell line classification processing unit that classifies the parenchymal cell region in units of branches, and the branch unit classified by the cell cord classification processing unit. For the pathological tissue image of the parenchymal cell region, the computer functions as a quantification processing unit including a distance measurement processing unit that measures the distance from the center line to the boundary of the parenchymal cell region and outputs the quantified pathological feature amount It can be made to.

  Furthermore, in the pathological tissue image analysis program according to the present invention, the quantification processing unit may further calculate and output the nucleus density for each cell cord or the average of the NC ratio for each cord as the pathological feature amount. it can.

  In the present invention, the region other than the blank region and the interstitial region is extracted from the pathological tissue image obtained by imaging the analysis target by extracting the region where the cells do not exist and the interstitial region in the analysis target region. To present the quantified information so that it can be easily used in clinical diagnosis by extracting as a parenchymal cell area in the analysis target area and quantifying the pathological feature amount for each aggregate of the parenchymal cells extracted as the parenchymal cell area. A pathological tissue image analysis method, a pathological tissue image analysis apparatus, and a pathological tissue image analysis program can be provided, and diagnosis support for pathological images can be realized.

  In particular, in the present invention, by analyzing and quantifying a cord-like structure that is a complex structure in an HE-stained pathological image of the liver, the quantified information can be presented so as to be easily used in clinical diagnosis. Can achieve the purpose.

It is a flowchart which shows the procedure of the pathological tissue image analysis method which concerns on this invention. It is a block diagram which shows the structure of the pathological tissue image analysis apparatus which implements the said pathological tissue image analysis method. It is a block diagram which shows the function of the computer which comprises the arithmetic processing part in the said pathological tissue image analyzer. It is a figure which shows the mode of a cellular cord (a color drawing is submitted as a reference figure). It is a flowchart which shows the processing content in the blank area | region extraction process step performed in the sinusoidal area | region extraction process part of the said pathological tissue image analysis apparatus. It is a figure which shows the mode of the boundary emphasis process by the azimuth | direction selection filter in the blank area | region extraction process step performed in the said sinusoid area | region extraction process part (a color drawing is submitted as a reference figure). It is a flowchart which shows the procedure of the erroneous extraction area | region reduction process in the said pathological tissue image analysis apparatus. It is a figure which shows a sinusoid area extraction result (a color drawing is submitted as a reference figure). It is a flowchart which shows the procedure of the color information extraction process in the said pathological tissue image analyzer. It is a figure which shows the area | region which calculates a color feature-value. It is a figure which shows the image used for learning. It is a flowchart which shows the processing content in the stromal area | region extraction process step performed in the stromal area | region extraction process part of the said pathological tissue image analysis apparatus. It is a figure which shows the state of the fiber and cytoplasm in a HE dyeing | staining image (a color drawing is submitted as a reference figure). It is a figure which shows the result of having extracted the fiber area | region using SVM (a color drawing is submitted as a reference figure). It is a figure which shows the result of superimposing the fiber probability calculated | required for every pixel on the super pixel (a color drawing is submitted as a reference figure). It is a figure which shows the result of having overlapped the fiber probability calculated | required for every pixel on the super pixel with many lymphocytes (a color drawing is submitted as a reference figure). It is a figure which shows the result of having extracted the super pixel with a high fiber probability and the super pixel with many lymphocytes as a stroma (submit a color drawing as a reference figure). It is a figure which shows the result of extracting a cell cord by extracting a sinusoid and stroma, and taking out the remainder (a color drawing is submitted as a reference figure). It is a flowchart which shows the processing content in the quantification process step performed in the quantification process part of the said pathological tissue image analyzer. It is a figure which shows the result of having piled up the branch extracted from the cord-like structure (a color drawing is submitted as a reference figure). It is a figure which shows the relationship between a cord-like structure and a branch (a color drawing is submitted as a reference figure). In the above pathological tissue image analyzer, the distance from the center line of the cord-like structure to the boundary is measured, and the result of automatically measuring the thickness of the cord by using the median as a representative value is compared with the result of manual measurement. It is the figure shown (a color drawing is submitted as a reference drawing). It is a figure which shows the histogram from which the thickness of a rope changes with typical differentiation degrees. It is a figure which shows a mode that it analyzes by the extracted branch unit (a color drawing is submitted as a reference figure). It is a figure which shows a mode that a cord unit is extracted from a cord-like structure. It is a figure which shows the example of classification | category by the differentiation degree of hepatocellular carcinoma. It is a figure which shows a nuclear density by setting a horizontal axis as a differentiation degree. It is a figure which shows the nuclear density average for every cord by making a horizontal axis into a differentiation degree. It is a figure which shows NC ratio by making a horizontal axis a differentiation degree. It is a figure which shows NC ratio average for every cord by making a horizontal axis into a differentiation degree. It is a figure which shows dispersion | distribution of the number of nuclei for every cord by making a horizontal axis into a differentiation degree. It is a figure which shows NC ratio average of the number of layers by making a horizontal axis into a differentiation degree. It is a figure (a color drawing is submitted as a reference figure) which shows the mode of the GUI screen which measures the thickness of a rope with the said pathological tissue image analyzer, and displays it on a display part. It is a figure which shows the image which expanded the pathological tissue image displayed on the [pathological tissue image] window in the said GUI screen (a color drawing is submitted as a reference figure). It is a figure which shows a liver pathological image.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Needless to say, the present invention is not limited to the following examples, and can be arbitrarily changed without departing from the gist of the present invention.

  The pathological tissue image analysis method according to the present invention includes, for example, an image capture processing step S1, a blank region extraction processing step S2, a stroma region extraction processing step S3, and a parenchymal cell region extraction processing step as shown in the flowchart of FIG. The pathological tissue image analysis apparatus 100 having a configuration as shown in FIG. 2, for example, includes a computer having S4 and quantification processing step S5 and executing the processing of each processing step S1 to S5.

  In the pathological tissue image analysis method according to the present invention, first, in the image capture processing step S1, a pathological tissue image obtained by imaging the analysis target is captured.

  In the next blank area extraction processing step S2, an area in which no cell exists in the analysis target area is extracted from the captured pathological tissue image.

  In the interstitial region extraction processing step S3, the interstitial region in the analysis target region is extracted from the pathological tissue image.

  In the next parenchymal cell region extraction processing step S4, the extracted regions other than the blank region and the stroma region are extracted from the pathological tissue image as the parenchymal cell region in the analysis target region.

  That is, in the pathological tissue image analysis method according to the present invention, a blank region and a stromal region are extracted from a pathological tissue image, and regions other than the blank region and the stromal region are extracted as a substantial cell region in the analysis target region.

  Then, in the quantification processing step S5, the pathological feature amount is quantified for each aggregate of parenchymal cells extracted as the parenchymal cell region.

  In the quantification processing step S5, the pathological feature quantity can be quantified as a group of the parenchymal cells in units of a series of cells having a certain length or less. Further, in the quantification processing step S5, the pathological feature quantity can be quantified based on a unit with a branch of a series of cells as a boundary as the assembly of the parenchymal cells.

  Further, as shown in the flowchart of FIG. 1, the pathological tissue image analysis method according to the present invention further includes a display processing step S6 for displaying the pathological feature value quantified in the quantification processing step S5. be able to.

  In the display processing step S6, for example, an index or a label representing the state of a lesion is displayed using a pathological feature value digitized for each aggregate of the parenchymal cells.

  Furthermore, the pathological tissue image analysis method according to the present invention includes an attention area mask generation step for generating mask information representing an area to be noted in the pathological tissue image. In the display processing step S6, the parenchymal cell is generated. The statistical value of the feature value in the attention area mask can be displayed using the pathological feature value digitized for each aggregate.

  Furthermore, the pathological tissue image analysis method according to the present invention includes a storing step of storing the pathological feature value quantified for each assembly of the parenchymal cells together with the label information assigned to each assembly. Can do.

  A pathological tissue image analysis apparatus 100 that performs a pathological tissue image analysis method according to the present invention includes an arithmetic processing unit 20 including a computer to which an image capture processing unit 10 is connected, and a main storage unit connected to the arithmetic processing unit 20. 30, an input unit 40, a display unit 50, and the like.

  In the histopathological image analysis apparatus 100, the image capture processing unit 10 includes a scanner and captures a pathological tissue image obtained by imaging an analysis target.

  The arithmetic processing unit 20 includes a computer that performs statistical processing for obtaining a pathological tissue image, processing the image, calculating an average of various feature amounts, a variance, a percentile, and the like.

  The main storage unit 30 stores a pathological tissue image analysis program and measurement results installed in the computer.

  The input unit 40 is composed of a keyboard and a mouse for designating a cord whose thickness is to be measured and a statistic calculation range.

  The display unit 50 displays the processing result by the arithmetic processing unit 20 and the like.

  Here, in the pathological tissue image analysis method according to the present invention, for example, in the image capture processing step S1, an image of a hepatocyte that has been subjected to a staining process is captured as the pathological tissue image, and the pathological tissue image of the liver cell is captured. Analysis can be performed.

  In this case, in the blank region extraction processing step S2, an sinusoidal region in the analysis target region is extracted from the captured pathological tissue image, and in the parenchymal cell region extraction processing step S4, the extracted sinusoidal region and the space between the extracted A region other than a quality region is extracted from the pathological tissue image as a substantial cell region in the analysis target region, and in the quantification processing step S5, a pathological feature amount is quantified for each cell line extracted as the substantial cell region. .

  Hereinafter, the configuration and operation of the pathological tissue image analysis apparatus 100 will be described in detail, assuming that the pathological tissue image analysis apparatus 100 analyzes a pathological tissue image of hepatocytes.

  In the pathological tissue image analysis apparatus 100, the computer constituting the arithmetic processing unit 20 executes the installed pathological tissue image analysis program, thereby capturing by the image capturing processing unit 10 as shown in FIG. A sinusoidal region extraction processing unit 210 that extracts a sinusoidal region in the analysis target region from the pathological tissue image obtained, and a stroma region in the analysis target region is extracted from the pathological tissue image captured by the image capture processing unit 10 The interstitial region extraction processing unit 220, the extracted sinusoidal region extracted by the sinusoidal region extraction processing unit 210, and the region other than the interstitial region extracted by the interstitial region extraction processing unit 220 are analyzed. A cell cord extraction processing unit 230 that extracts from the pathological tissue image as a cell cord region in the target region, and the cell cord extraction processing unit 230 A quantification processing unit 240 that quantifies a pathological feature amount for each cell line extracted as a cell cord region, and a display control unit 250 that performs control to display a processing result by the quantification processing unit 240 on the display unit 50. Function as.

  That is, the arithmetic processing unit 20 in the pathological tissue image analysis apparatus 100 extracts an sinusoidal region in the analysis target region from the pathological tissue image captured by the image capture processing unit 10 as the processing of the blank region extraction processing step S2. The sinusoidal region extraction processing unit 210 that performs processing, and the interstitial region extraction processing step S3 that extracts the interstitial region in the analysis target region from the pathological tissue image captured by the image capture processing unit 10 are performed. The interstitial region extraction processing unit 220, the extracted sinusoidal region extracted by the above-mentioned interstitial region extraction processing unit 210, and regions other than the interstitial region extracted by the above-mentioned interstitial region extraction processing unit 220 are analyzed. Cell Rout Extraction for Performing the Process of the Parenchymal Cell Region Extraction Processing Step S4 for Extracting from the Pathological Tissue Image as a Real Cell Region in the Region It includes a processing section 230, and a quantifying unit 240 for processing the quantification process step S5 to quantify pathological feature amount for each cell cord extracted as substantially cell areas by the cell search extraction processing section 230.

  Here, in the present embodiment, the pathological tissue image is, for example, an image of a hepatocyte subjected to a staining process.

  Since hepatocellular carcinoma is based on a cord-like structure that mimics the structure of normal liver, it is necessary to analyze the cord-like structure to distinguish between cells that show normal structure and cancer cells that mimic the cord-like structure .

  The state of the cell cord 1 is shown in FIG.

  The sinusoidal wall cells such as endothelial cells that exist at the boundary between the sinusoid and hepatocytes can be recognized visually as arranged on a smooth curve, but there are gaps and gaps due to the influence of red blood cells and lymphocytes. If an area is divided based on this boundary by image processing, it is difficult to obtain a correct boundary at the interrupted portion.

  Therefore, in this pathological tissue image analysis apparatus 100, an orientation selection filter is applied to connect a discontinuous portion to a discontinuous boundary having a smooth shape and emphasize the boundary.

  In the processing of the blank region extraction processing step S2 executed in the sinusoidal region extraction processing unit 210, that is, the sinusoidal region extraction processing, as shown in the flowchart of FIG. Boundary enhancement processing is performed on the image using an orientation selection filter (boundary enhancement processing step S21).

  Here, the azimuth selection filter selectively convolves a bar-shaped filter that is affine-transformed in the gradient direction of luminance. FIG. 6B shows the result of applying the orientation selection filter in the boundary enhancement processing step S21 to the original image shown in FIG.

  In the original image shown in FIG. 6 (A), the cytoplasm around the sinusoids is lightened and the boundary is unclear when enlarged. On the other hand, in the result of applying the orientation selection filter shown in FIG. 6B, the boundary with the lighted cytoplasm becomes darker in the pink color of eosin due to the surrounding cytoplasm, and the sinusoids are secreted internally. It can be confirmed that the liquid is whitened by the smoothing action, and the boundary between the sinusoids and the hepatocytes is clear.

  Then, the sinusoidal region extraction processing by clustering is performed on the pathological tissue image of the analysis target region subjected to the boundary enhancement processing in the boundary enhancement processing step S21 (clustering processing step S22). In this clustering processing step S22, a white region is extracted as a sinusoid by performing clustering by applying an EM algorithm to the processing result image.

  Further, an erroneous extraction region reduction process using a support vector machine (SVM) is performed on the sinusoidal region extracted in the clustering processing step S22 (error extraction region reduction processing step S23).

  In the histopathological image analysis apparatus 100, the erroneous extraction region reduction processing in the erroneous extraction region extraction reduction processing step S23 is performed in accordance with the procedure shown in the flowchart of FIG. 7, color information extraction processing step S231, learning processing step S232, and deletion processing step S233. Is done.

  Here, FIG. 8B shows a result extracted from the original image shown in FIG. 8A using the azimuth selection filter and the EM algorithm. In FIG. 8B, mis-extraction is seen in the lightened cytoplasm. Therefore, here, the small region extracted in FIG. 8B is referred to as a sinusoidal candidate region. Then, an erroneous extraction area reduction process is performed to improve extraction accuracy. The final reduction result is shown in FIG.

  In color information extraction processing step S231, histogram creation processing step S231A, outer region determination processing step S231B, feature amount calculation are performed for each small region extracted as shown in FIG. 8B according to the procedure shown in the flowchart of FIG. Color information is extracted by processing step S231C.

  In the histogram creation processing step S231A, a 64-gradation histogram is created for each extracted small region.

  In the outer area determination processing step S231B, each of the small areas is expanded five times, and the expanded area is defined as the outer area, thereby determining an outer area corresponding to the outer 5 pixels of the target small area 2.

  In the feature amount calculation processing step S231C, a histogram feature amount is calculated in the small region 2 extracted as shown in FIG. 10A based on the histogram created in the histogram creation processing step S231A. Further, as shown in FIG. 10B, the color feature amount of the outer region 3 for 5 pixels outside the target small region is calculated.

As the color feature amount, for example, an average (μ), a variance (σ ² ), a skewness (S), and a kurtosis (K) are obtained from the density histogram by the following formulas (1) to (4).

  In the learning processing step S232, the SVM is learned with the color feature values of the “analogue and airspace other than the sinusoids” using the 8-dimensional feature values extracted in each small region in this way. The images shown in FIGS. 11A and 11B were used for learning. In the images shown in FIGS. 11A and 11B, green indicates the sinusoid 5, red indicates the airspace 6 other than the sinusoid, and 173 airspaces corresponding to the sinusoid, other than the sinusoid There are 108 airspaces.

  In the deletion processing step S233, when a new image is input, a sinusoidal candidate region is similarly extracted using an orientation selection filter and an EM algorithm, and a support vector machine (SVM: Support Vector Machine) learned by the above-described method is used. ) To reduce erroneous extraction areas.

  As described above, the sinusoidal region extraction processing unit 210 performs boundary enhancement processing using the orientation selection filter for the pathological tissue image of the analysis target region, and performs clustering for the pathological tissue image of the analysis target region subjected to the boundary enhancement processing. The sinusoidal region can be extracted accurately by performing the extraction region reduction process using a support vector machine (SVM) on the extracted sinusoidal region.

  That is, in the pathological tissue image analysis method according to the present invention, the sinusoidal region extraction processing step S2 performs boundary enhancement processing using a directional selection filter on the pathological tissue image of the analysis target region, as shown in FIG. Processing step S21, clustering processing step S22 for performing the sinusoidal region extraction processing by clustering on the pathological tissue image of the analysis target region subjected to the boundary enhancement processing, and the support vector machine (SVM) for the extracted sinusoidal region : Support Vector Machine), it is possible to accurately extract a sinusoid by including an erroneous extraction area reduction processing step S23 for performing an erroneous extraction area reduction process.

  In the processing of the interstitial region extraction processing step S3 executed in the interstitial region extraction processing unit 220, as shown in FIG. 12, the pathological tissue image of the analysis target region is processed in the fiber probability calculation processing step S31. In addition to calculating the fiber probability based on the texture features, in the superpixel processing step S32, superpixel processing is performed to cluster the pathological tissue images in the analysis target region into regions having similar colors, and based on the processing results. Then, in the first interstitial extraction processing step S331, an average fiber probability is calculated for each superpixel and a superpixel having a high fiber probability is extracted as a stroma, and in the second interstitial extraction processing step S332, lymphocytes are extracted. Extract super-pixels with a lot of interstitial material, Interstitial area extracting process for extracting a interstitial a result of combined high superpixel of lymphocytes (interstitial extraction process step S33) performed.

  That is, in HE staining, cell nuclei are stained with violet hematoxylin and cytoplasm with violet eosin. At that time, eosin also enters the fiber, so it stains reddish purple near the cytoplasm. For this reason, extraction is difficult by clustering based on color like a sinusoid. Fibers are shown in FIG. 13 (A) (see color drawing submitted as a reference diagram), and cytoplasm is shown in FIG. 13 (B) (see color drawing submitted as a reference drawing). In addition, in the HE-stained image, the color difference between the fiber and the cytoplasm is fine, and the color density changes relatively depending on the staining state, so that extraction by color information has many restrictions.

  Therefore, in this pathological tissue image analysis apparatus 100, identification by texture feature is performed using a density co-occurrence matrix conventionally used in medical image processing, and fiber probability by texture feature is calculated for each pixel in fiber probability calculation processing step S31. To do.

  The density co-occurrence matrix is a technique for quantifying the co-occurrence between pixels within a certain range of rectangles, and is especially effective when the texture has directionality because it is quantified in consideration of directionality. To do. As shown in FIG. 13A, the fiber often has a directional texture. In contrast, as shown in FIG. 13B, the cytoplasm and the cell nucleus are isotropic, and the texture characteristics do not change greatly depending on the orientation. Therefore, 2000 fiber regions and 2000 regions of 64 × 64 size were collected, and learning was performed using a support vector machine (SVM).

  Here, a 1 mm × 1 mm (2174 × 2174 pixel) image photographed at 20 × is targeted, and the result of extracting the fiber region from the image shown in FIG. 14A using SVM is shown in FIG. ). It can be confirmed that the fiber is extracted at the portion having directionality.

  In the extraction result shown in FIG. 14B, since the calculation is performed in the range of 64 × 64, the resolution is low with respect to the original image. In addition, since lymphocytes other than fibers enter the stroma, it is necessary to add a lymphocyte region.

  In the histopathological image analysis apparatus 100, superpixels are introduced to solve two problems. Superpixel is a method of clustering in a well-colored region, and methods using mean shift method, k-means method, and graph cut have been proposed. In the super pixel processing step S32, a super pixel is calculated using SLIC (Simple Linear Iterative Clustering) which is conventionally known. For example, SLIC is described in detail in Non-Patent Document 3.

  In the stromal region extraction processing step S33 in the pathological tissue image analysis apparatus 100, since the fiber probability is obtained for each pixel in the fiber probability calculation processing step S31, in the first stromal extraction processing step S331, An average fiber probability is calculated for each superpixel, and a superpixel with a high fiber probability is extracted as a stroma, and a superpixel with many lymphocytes is extracted as a stroma in the second stroma extraction processing step S332, and a fiber probability is extracted. And the result of combining superpixels with many lymphocytes is extracted as stroma.

  An example of the processing result of the interstitial region extraction processing step S3 is shown in FIGS.

  FIG. 15B shows the result of superimposing the fiber probability shown in FIG. 15A obtained for each pixel in the fiber probability calculation processing step S31 on the superpixel in the first stroma calculation processing step S331.

  Also, FIG. 16B shows the result of superimposing the fiber probability shown in FIG. 16A obtained for each pixel in the fiber probability calculation processing step S31 on the superpixel in the second stroma calculation processing step S331. Show.

  Then, from the fiber probability shown in FIG. 17A obtained for each pixel in the fiber probability calculation processing step S31, the superpixel having a high fiber probability extracted in the first interstitial extraction processing step S331 and the second pixel. FIG. 17B shows the result of extracting the fiber probability extracted in the quality extraction processing step S332 and the superpixel having many lymphocytes as the stroma.

  From FIG. 17B, it can be confirmed that the interstitial region is extracted with high accuracy.

  As described above, in the interstitial region extraction processing step S3, in the first interstitial extraction processing step S331, the average fiber probability is calculated for each superpixel, and a superpixel having a high fiber probability is extracted as the stroma. 2 In the interstitial extraction processing step S332, superpixels with many lymphocytes are extracted as stroma, and the result of combining the fiber probability and the superpixel with many lymphocytes is extracted as stroma, so that the stromal region is accurately extracted. Can be extracted.

  In the present embodiment, SLIC is adopted as a method for calculating a super pixel, but the present invention is not limited to this.

  Then, in the histopathological image analysis apparatus 100, the cell line extraction processing unit 230 executes the processing of the parenchymal cell region extraction processing step S4, thereby extracting the sinusoids and stroma and taking out the rest. Extract the cell cord. The extracted result is shown in FIG. FIG. 18 is enlarged for easy viewing of the extraction result. In FIG. 18, green indicates the stroma 7 and blue indicates the sinusoid 8. It can be confirmed that the cell cord has been successfully extracted.

  Further, in the histopathological image analysis apparatus 100, the quantification processing unit 240 performs the quantification processing step S5 on the cell cord extracted by the cell cord extraction processing unit 230 as a parenchymal cell region excluding the sinusoid and the stroma. This process is executed according to the procedure shown in the flowchart of FIG.

  The quantification processing step S5 in the histopathological image analysis apparatus 100 includes a thinning processing step S51, a cell cord classification processing step S52, and a distance measurement processing step S53, and is extracted in the parenchymal cell region extraction processing step S4. With respect to the pathological tissue image of the cell line area, the distance along the center line of the closed area figure of the cell line is measured, and distance measurement processing is performed to output as a quantified pathological feature amount.

  That is, in the thinning processing step S51, the parenchymal cell region is thinned by obtaining the center line of the parenchymal cell region extracted in the parenchymal cell region extraction processing step S4, and the thinning is performed in the cell cord classification processing step S52. The parenchymal cell region is classified into branches from branch to branch of the parenchymal cell region. In the distance measurement processing step S53, the distance from the center line of the cell cord region to the boundary is measured and quantified with respect to the pathological tissue image of the cell unit region of the branch unit classified in the cell cord classification processing step S52. Is output as the pathological feature value.

  Further, in the quantification processing step S5, it is possible to further calculate and output the nucleus density for each cell cord or the average of the NC ratio for each cord as the pathological feature amount.

  Here, the cord-like structure is continuous like a chain, but the cell structure such as the thickness of the cord and the NC ratio varies depending on the location within the 1 mm × 1 mm visual field. Therefore, this pathological tissue image analysis apparatus 100 employs a method of performing quantification after performing preprocessing for dividing the cord-like structure into a certain length. Specifically, it is quantified by labeling with a branch unit of the center line obtained by thinning the cord-like region. The result of overlapping branches extracted from the cord-like structure is shown in FIG. In FIG. 20, the green line indicates the center line 9 of the cord-like structure.

  The center line 9 is obtained using a distance conversion image from a sinusoid or a stroma. In this process, since each pixel of the thinned image knows the distance to the boundary, the pathological tissue image analysis apparatus 100 automatically measures the thickness of each cord.

  Here, FIG. 21 shows a specific relationship between the cord-like structure and the branches. FIG. 21B shows the distance between each pixel in the branch from point A to point B in the image shown in FIG.

  Here, it can be seen that the distance tends to increase near the branch due to the algorithm problem of the distance conversion image. Therefore, in the histopathological image analysis apparatus 100, the distance from the center line to the boundary is measured from the point C to the point D excluding only the thickness of the branch from the branch, and the median value is used as the representative value. Automatic thickness measurement.

  An actual measurement example is shown in FIG. In FIG. 22, a portion indicated by a black line is manually measured and compared with a result of automatic measurement by the pathological tissue image analysis apparatus 100. Although a shift of about 2 to 3 μm occurs depending on the manual measurement and the measurement location, it can be confirmed that the result is quite close compared to the measurement location. Conventionally, many computer diagnosis supports display with the number of pixels and ratio, and it has been difficult for a doctor to intuitively understand the feature quantity itself. However, in the pathological tissue image analysis apparatus 100, the measurement result can be displayed not only as a feature amount but also as information that can be intuitively utilized in diagnosis by indicating the measurement result in μm.

  Specifically, first, in FIG. 21A, the distance to the closest sinusoid or interstitium in all the pixels on the center line from the point A to the point B where the thickness is measured is calculated. This corresponds to the radius of the cell cord. For this reason, the thickness of the cell cord is defined as twice the radius. Strictly speaking, there is a possibility that the distance from the center line to the boundary may be different between the left and right sides, but it is considered that there is no significant influence because an error occurs even in manual measurement. The thickness of one cell cord is the median value when the thicknesses of all the pixels are measured. This is to reduce the influence of outliers due to errors in the extraction of sinusoids and stroma. In addition, since the calculation is inaccurate near the branch point corresponding to the orange range in FIG. 21B, the thickness of the target cell cord from the branch point is not included in the calculation.

  Further, the thickness of the cord tends to increase as the degree of differentiation decreases, and only the thickness of the cord is displayed as a histogram, which is one piece of information for estimating the degree of differentiation. FIG. 23 shows a histogram in which the thickness of the cord varies with a typical degree of differentiation. As shown in FIG. 23, it can be seen that the thickness of the cords varies and the number of thick cords increases as the degree of differentiation decreases.

  Furthermore, the analysis is performed for each extracted branch. This is equivalent to structural analysis for each cord. As shown in FIG. 24, the cord unit performs the analysis from the branch of the cord to the branch as one region. In addition, as shown by arrows in FIG. 24, a cord label is assigned to the nuclear cord. The cable structure is extracted from the branch using a convex hull as shown in FIG. FIG. 25 shows a state where the cord label S-003 of FIG. 24 is extracted.

  That is, as shown in FIG. 25 (A), the branches of the thinning result are expanded, the convex hull is calculated, and the cellular cords that enter the convex hull from the cellular cord structure shown in FIG. Remove as shown in (C). At this time, since other regions may be included in the convex hull, only the region with the largest area is extracted by labeling. However, the method for determining the area of the rope is not limited to this method.

  In this pathological tissue image analysis apparatus 100, the convex hull region shown in FIG. 25C is considered as one cell cord, and the NC ratio, nuclear density, curvature, and number of layers are obtained.

  The NC ratio is generally obtained by dividing the cell nucleus area by the cytoplasm area within the visual field range (1 mm × 1 mm or the like). Actually, in clinical practice, since it is difficult to accurately detect cell nuclei and cytoplasm, it is often required by eye measurement.

  In the histopathological image analysis apparatus 100, the cytoplasm is accurately obtained as described above, and is obtained with high accuracy by the cell nucleus detection method described in Non-Patent Document 4, for example. In Non-Patent Document 4, hematoxylin and eosin (HE) -stained histopathological specimen images of liver are treated with hematoxylin color (blue purple) that stains nuclei in RGB color space and eosin (pink) that stains cytoplasm. ), The RGB signal values are projected, the contrast between the cell nucleus and the cytoplasm is improved, and finally the cell nucleus is detected using template matching. The method for detecting cell nuclei is not limited to the method for detecting cell nuclei described in Non-Patent Document 4, and other methods such as the method for detecting cell nuclei described in Non-Patent Document 5 may be employed. it can.

  Moreover, the result of overlapping the cell nucleus on the cell cord is shown in FIG. In FIG. 25D, a region P shown in gray is a cell nucleus. Therefore, the NC ratio for each cord label can be obtained by the area of the cell nucleus (number of gray pixels) ÷ the area of the cytoplasm (number of white pixels).

  Nuclear density means the ratio of the number of cell nuclei within a certain range. Usually, it is calculated by the number of cell nuclei existing within the visual field range (1 mm × 1 mm) ÷ the area of cytoplasm as in the NC ratio.

  In the histopathological image analysis apparatus 100, as shown in FIG. 25D, it can be obtained by the number of gray regions P ÷ the area of cytoplasm (the number of white pixels).

  It is also possible to obtain the number of layers of each cell cord as a feature amount closer to a pathological finding.

In the histopathological image analysis apparatus 100, the length and thickness of one cord and the number of contained cell cords are obtained. First, the area (a ² ) of one cell nucleus can be obtained by the following equation (5) when a cell is approximated by a square.

  At this time, t is the thickness of the cell cord, l is the length of the cell cord, and m is the number of cell nuclei contained in the cell cord. Therefore, the number of layers of cell cords is defined as the following equation (6). a is the diameter when one cell is regarded as a square.

  In addition, although the medical relationship is not clear, the curvature can be obtained from the center line of the obtained cell cord. The curvature is generally defined as the reciprocal of the radius of curvature, but can be approximated by changing the angle of the tangent vector in the discrete space.

In the histopathological image analysis apparatus 100, cell cords are extracted and the thickness of the cell cord is measured in units of μm, so that it can be acquired and displayed as information that can be intuitively used in pathological diagnosis. In addition, when quantification / counting is performed on a HE-stained pathological specimen image, the structure is complex, and therefore, the calculation is often performed within a certain range by dividing it with a rectangle or the like. Since the region is divided into units and quantified, it can be analyzed by a method suitable for a more pathological diagnosis policy.
[Example 1]

  The object to be analyzed by the pathological tissue image analyzing apparatus 100 is a cord-like structure, and 29 pathological tissue images having different degrees of differentiation were collected for experiments. There are 10 background livers, 6 highly differentiated, 4 high to moderately differentiated, and 9 moderately differentiated. In this Example 1, the compared data are the number of nuclei and the NC ratio.

  When diagnosing hepatocellular carcinoma, the degree of differentiation classified from cell / structure variants is used. The degree of differentiation is classified into well differentiated hepatocellular carcinoma, moderately differentiated hepatocellular carcinoma, and poorly differentiated hepatocellular carcinoma. The closer to poorly differentiated cancer, the stronger the cell / structure variants. In addition, as shown in FIG. 26, the degree of differentiation is not simply composed of only the moderately differentiated type within one field of view but is determined in detail by looking at the tissue structure.

  In FIG. 26, I, II, and III indicate a highly differentiated type, a moderately differentiated type, and a poorly differentiated type, respectively. It is described as moderately differentiated hepatocellular carcinoma with poorly differentiated cancer tissue.

  It is known that the closer to the poorly differentiated cancer, the more accompanied the cell atypia such as nuclear enlargement and increase in NC ratio (nucleus ÷ cytoplasm). Table 1 shows the results of the correlation analysis for the number of nuclei in the entire image, the NC ratio, and each cord.

  Since the degree of differentiation is not understood by only one evaluation factor, the correlation coefficient is low, but the average of the number of nuclei contained in each cell cord is less than the number of simple nuclei within 1 mm × 1 mm. The correlation coefficient is higher. Similarly, the correlation coefficient of the NC ratio is higher when the NC ratio is obtained for each cell cord and averaged than the ratio of cell nucleus / cytoplasm in the entire image. This is because, as shown in FIG. 26, a region with strong cell atypia and a region with weak cell affinities are mixed in one image. It is thought that it is shown. FIG. 27 shows the nuclear density with the horizontal axis as the degree of differentiation, and FIG. 28 shows the average nuclear density for each cord with the horizontal axis as the degree of differentiation. FIG. 29 shows the NC ratio with the horizontal axis as the degree of differentiation, and FIG. 30 shows the NC ratio average for each cord with the horizontal axis as the degree of differentiation.

  Another advantage of evaluating pathological images for each cord is that structural analysis can be performed in units of cords. For example, looking at the degree of differentiation and the distribution of the number of nuclei for each cord, the correlation coefficient is as high as 0.77.

  FIG. 31 shows the distribution of the number of nuclei for each cord with the horizontal axis as the degree of differentiation. FIG. 31 shows that the number of nuclei entering one cord varies as the degree of differentiation decreases.

Further, FIG. 32 shows the average number of cell cord layers for each cord with the horizontal axis as the degree of differentiation. As shown in Table 1, the average number of layers shows a high correlation coefficient of 0.74. Regarding the calculation of the number of layers, it is significant that the number of layers can be clearly defined for each line. In addition, quantitative analysis is facilitated by dividing cell cords that are continuous in the form of cords into cord labels.
[Example 2]

  In the pathological tissue image analysis apparatus 100, information for each cord label as shown in Table 2 is stored in the main storage unit 30, and a pathology obtained by imaging an analysis target according to a pathological tissue image analysis program. The tissue image was captured by the image capture processing unit 10, the captured pathological tissue image was automatically measured by the arithmetic processing unit 20, and the measurement result was stored in the main storage unit 30.

  FIG. 33 shows a state in which the thickness of the cord is measured by the pathological tissue image analysis apparatus 100 and displayed on the display unit 50.

  In the above-mentioned pathological tissue image analysis apparatus 100, when the [Measure Thickness] button 50A is clicked on the GUI screen shown in FIG. 33 displayed on the display unit 50, the measurement mode is entered, and the center line of the cord whose thickness is to be known is clicked. Then, the corresponding center line 9 is displayed in green, and the thickness of the cord is displayed in yellow letters. FIG. 34 shows an enlarged image of the pathological tissue image displayed in the [pathological tissue image] window 50B.

  Next, in the GUI screen shown in FIG. 33, in the [Image Information Display] window 50C, the above-described nuclear density, the thickness of the entire cord, the average nuclear density for each cord, the curvature of the cord, the number of layers of the cellular cord, etc. It can be displayed and compared.

  In the GUI screen shown in FIG. 33, a histogram of the thickness of the cell cord can be confirmed in the [Histogram display] window 50D.

  DESCRIPTION OF SYMBOLS 10 Image acquisition process part, 20 Arithmetic process part, 30 Main memory part, 40 Input part, 50 Display part, 100 Pathological tissue image analyzer, 210 Sinusoidal area extraction process part, 220 Stromal area extraction process part, 230 cell Rope extraction processing unit, 240 quantification processing unit, 250 display control unit

Claims (42)

An image capture processing step for capturing a pathological tissue image obtained by imaging the analysis object;
A blank area extraction processing step for extracting an area where cells do not exist in the analysis target area from the captured pathological tissue image,
A stroma region extraction processing step for extracting a stroma region in the analysis target region from the pathological tissue image;
A parenchymal cell region extraction processing step for extracting a region other than the extracted blank region and the stroma region from the pathological tissue image as a parenchymal cell region in the analysis target region;
A pathological tissue image analysis method comprising: a quantification processing step for quantifying a pathological feature amount for each aggregate of parenchymal cells extracted as the parenchymal cell region.   The pathological tissue image analysis method according to claim 1, further comprising a display step of displaying the pathological feature value quantified in the quantification processing step.   3. The pathological tissue image analysis method according to claim 2, wherein, in the display step, an index or a label representing a lesion state is displayed using a pathological feature value digitized for each aggregate of parenchymal cells. . Furthermore, it has an attention area mask generation step for generating mask information representing an area to be noted in the pathological tissue image,
3. The pathological tissue according to claim 2, wherein, in the display step, a statistical value of the feature amount in the attention area mask is displayed using a pathological feature amount digitized for each aggregate of the parenchymal cells. Image analysis method.   2. The pathological tissue image analysis method according to claim 1, wherein, in the quantification processing step, the pathological feature amount is quantified as a group of the parenchymal cells in units of a series of cells having a predetermined length or less.   2. The pathological tissue image analysis method according to claim 1, wherein, in the quantification processing step, the pathological feature amount is quantified based on a unit having a boundary of a series of cells as a boundary as the aggregate of the parenchymal cells. .   The pathological tissue image analysis method according to claim 1, further comprising a storage step of storing the pathological feature value digitized for each aggregate of parenchymal cells together with label information assigned to each aggregate.   The pathological tissue image analysis method according to any one of claims 1 to 7, wherein the histopathological image is an image of a hepatocyte subjected to a staining process. As the blank region extraction processing step, there is a sinusoidal region extraction processing step for extracting a sinusoidal region in the analysis target region from the captured pathological tissue image,
In the parenchymal cell region extraction processing step, a region other than the extracted sinusoidal region and the stroma region is extracted from the pathological tissue image as a parenchymal cell region in the analysis target region,
9. The pathological tissue image analysis method according to claim 8, wherein in the quantification processing step, a pathological feature amount is quantified for each cell line extracted as the parenchymal cell region. The above sinusoidal region extraction processing step includes:
Boundary enhancement processing step for performing boundary enhancement processing with an orientation selection filter for the pathological tissue image of the analysis target region,
A clustering processing step of performing an sinusoidal region extraction process by clustering on the pathological tissue image of the analysis target region subjected to the boundary enhancement processing;
The pathological tissue image analysis method according to claim 9, further comprising: an erroneous extraction region reduction processing step of performing an erroneous extraction region reduction process by a support vector machine (SVM) for the extracted sinusoidal region. . The interstitial region extraction processing step includes
For the pathological tissue image of the analysis target region, a fiber probability calculation processing step for performing fiber probability calculation processing by texture features;
For the pathological tissue image of the region to be analyzed, a superpixel processing step for performing a superpixel processing for clustering into regions having similar colors,
A first interstitial extraction processing step of calculating an average fiber probability for each superpixel and extracting a superpixel having a high fiber probability as a stroma;
11. The histopathological image analysis method according to claim 9, further comprising a second stroma extraction processing step of extracting a superpixel having a large number of lymphocytes as a stroma.   In the quantification processing step, for the pathological tissue image of the parenchymal cell region extracted in the parenchymal cell region extraction processing step, the width along the center line of the closed region figure of the parenchymal cell region is measured and quantified pathology The pathological tissue image analysis method according to any one of claims 9 to 11, wherein a distance measurement process for outputting as a feature amount is performed. The quantification processing step is
For the pathological tissue image of the parenchymal cell region extracted in the parenchymal cell region extraction processing step, a thinning step for thinning by obtaining a center line of the parenchymal cell region;
A cell line classification processing step for classifying the parenchymal cell region in units of branches, with the branch from the branch to the branch of the thinned parenchymal cell region as a branch,
A distance measurement processing step of measuring the distance from the center line of the parenchymal cell region to the boundary for the pathological tissue image of the classified parenchymal cell region, and outputting the quantified pathological feature amount. The pathological tissue image analysis method according to claim 12.   14. The pathological tissue image analysis method according to claim 13, wherein in the quantification processing step, the average of the nuclear density for each cell line or the NC ratio for each line is further calculated and output as the pathological feature amount. An image capture processing unit that captures a pathological tissue image obtained by imaging the analysis object;
A blank region extraction processing unit that extracts a region where cells do not exist in the analysis target region from the pathological tissue image captured by the image capture processing unit,
A stroma region extraction processing unit that extracts a stroma region in the analysis target region from the pathological tissue image;
A parenchymal cell that extracts a blank region extracted by the stromal region extraction processing unit and a region other than the stromal region extracted by the stromal region extraction processing unit as a substantial cell region in the analysis target region from the pathological tissue image A region extraction processing unit;
A pathological tissue image analysis apparatus comprising: a quantification processing unit that quantifies a pathological feature amount for each assembly of parenchymal cells extracted as a parenchymal cell region by the parenchymal cell region extraction processing unit.   The pathological tissue image analysis apparatus according to claim 15, further comprising a display unit that displays the pathological feature value quantified by the quantification processing unit.   The said quantification process part displays the parameter | index or label showing the state of a lesion | pathology on the said display part using the pathological feature-value digitized for every aggregate | assembly of the said parenchymal cell, The said display part is characterized by the above-mentioned. Histopathological image analyzer.   The quantification processing unit generates mask information representing a region to be noted in the pathological tissue image, and uses the pathological feature value quantified for each aggregate of the parenchymal cells to generate a mask information in the region of interest mask. The pathological tissue image analysis apparatus according to claim 16, wherein the statistical value of the feature amount is displayed on the display unit.   The pathological tissue image analysis apparatus according to claim 15, wherein the quantification processing unit quantifies a pathological feature amount in a unit of a series of cells having a predetermined length or less as the aggregate of the parenchymal cells.   16. The pathological tissue image analysis apparatus according to claim 15, wherein the quantification processing unit quantifies the pathological feature amount based on a unit having a boundary of a series of cells as a boundary as the aggregate of the parenchymal cells. .   The pathological tissue image analysis according to claim 15, further comprising an information storage unit that stores the pathological feature value digitized for each aggregate of parenchymal cells together with label information assigned to each aggregate. apparatus.   The histopathological image analysis apparatus according to any one of claims 15 to 21, wherein the histopathological image is an image of a hepatocyte subjected to a staining process. The blank region extraction processing unit is an sinusoidal region extraction processing unit that extracts a sinusoidal region in the analysis target region from the captured pathological tissue image,
In the parenchymal cell region extraction processing unit, the region other than the sinusoidal region and the stroma region extracted by the sinusoidal region extraction processing unit is extracted from the pathological tissue image as a parenchymal cell region in the analysis target region,
23. The pathological tissue image analysis apparatus according to claim 22, wherein the quantification processing unit quantifies a pathological feature amount for each cell line extracted as a substantial cell region by the substantial cell region extraction processing unit. The sinusoidal region extraction processing unit is
A boundary emphasis processing unit that performs a boundary emphasis process using an orientation selection filter on the pathological tissue image of the analysis target region;
A clustering processing unit that performs a sinusoidal region extraction process by clustering on the pathological tissue image of the analysis target region that has been subjected to boundary enhancement processing by the boundary enhancement processing unit;
24. A false extraction region reduction processing step of performing false extraction region reduction processing by a support vector machine (SVM) for the sinusoidal region extracted by the clustering processing unit. Pathological tissue image analyzer. The interstitial region extraction processing unit
For the pathological tissue image of the analysis target region, a fiber probability calculation processing unit that performs fiber probability calculation processing by texture features;
With respect to the pathological tissue image of the analysis target region, a superpixel processing unit that performs superpixel processing for clustering into regions having similar colors,
A first stroma extraction processing unit that calculates an average fiber probability for each superpixel and extracts a superpixel having a high fiber probability as a stroma;
The pathological tissue image analysis apparatus according to any one of claims 23 and 24, further comprising: a second stroma extraction processing unit that extracts a superpixel having a large number of lymphocytes as a stroma.   In the quantification processing unit, the pathology tissue image of the parenchymal cell region extracted by the parenchymal cell region extraction processing unit is measured by measuring the width along the center line of the closed region figure of the parenchymal cell region, and quantified pathology The pathological tissue image analysis apparatus according to any one of claims 23 to 25, wherein a distance measurement process for outputting as a feature amount is performed. The quantification processing unit
For the pathological tissue image of the parenchymal cell region extracted by the parenchymal cell region extraction processing unit, a thinning processing unit for thinning by obtaining the center line of the parenchymal cell region,
A cell line classification processing unit that classifies the parenchymal cell region in units of branches, with the branch from the branch to the branch of the thinned parenchymal cell region as a branch,
Distance measurement for measuring the distance from the center line of the parenchymal cell region to the boundary and outputting it as a quantified pathological feature amount for the pathological tissue image of the parenchymal parenchymal cell region classified by the cell cord classification processing unit 27. The pathological tissue image analysis apparatus according to claim 26, further comprising: a processing unit.   28. The histopathological image analysis apparatus according to claim 27, wherein the quantification processing unit further calculates and outputs a nuclear density for each cell line or an average of an NC ratio for each line as a pathological feature amount. A pathological tissue image analysis program executed by a computer provided in a pathological tissue image analysis apparatus that analyzes a pathological tissue image obtained by imaging an analysis object,
A blank area extraction processing unit that extracts an area where cells do not exist in the analysis target area from the captured pathological tissue image;
A stroma region extraction processing unit that extracts a stroma region in the analysis target region from the pathological tissue image;
A parenchymal cell that extracts a blank region extracted by the stromal region extraction processing unit and a region other than the stromal region extracted by the stromal region extraction processing unit as a substantial cell region in the analysis target region from the pathological tissue image A region extraction processing unit;
A pathological image analysis program that causes the computer to function as a quantification processing unit that quantifies a pathological feature amount for each aggregate of parenchymal cells extracted as a parenchymal cell region by the parenchymal cell region extraction processing unit. .   30. The pathological image analysis program according to claim 29, wherein the pathological feature value quantified by the quantification processing unit is displayed on a display unit.   31. The quantification processing unit displays an index or a label representing a lesion state on the display unit by using a pathological feature value quantified for each aggregate of parenchymal cells. Pathological image analysis program.   The quantification processing unit generates mask information representing a region to be noted in the pathological tissue image, and uses the pathological feature value quantified for each aggregate of the parenchymal cells to generate a mask information in the region of interest mask. 31. The pathological image analysis program according to claim 30, wherein statistical values of feature quantities are displayed on the display unit.   30. The pathological image analysis program according to claim 29, wherein the quantification processing unit quantifies a pathological feature amount in units of a series of cells having a predetermined length or less as the aggregate of the parenchymal cells.   30. The pathological image analysis program according to claim 29, wherein the quantification processing unit quantifies a pathological feature amount based on a unit having a branch of a series of cells as a boundary as an aggregate of the parenchymal cells.   30. The pathological image analysis program according to claim 29, wherein the pathological feature value digitized for each aggregate of parenchymal cells is stored in an information storage unit together with label information given for each aggregate.   36. The pathological image analysis program according to claim 29, wherein the pathological tissue image is an image of a hepatocyte subjected to a staining process. The blank region extraction processing unit is an sinusoidal region extraction processing unit that extracts a sinusoidal region in the analysis target region from the captured pathological tissue image,
In the parenchymal cell region extraction processing unit, the region other than the sinusoidal region and the stroma region extracted by the sinusoidal region extraction processing unit is extracted from the pathological tissue image as a parenchymal cell region in the analysis target region,
37. The pathological image analysis program according to claim 36, wherein the quantification processing unit quantifies a pathological feature amount for each cell line extracted as a substantial cell region by the substantial cell region extraction processing unit. The sinusoidal region extraction processing unit is
A boundary emphasis processing unit that performs a boundary emphasis process using an orientation selection filter on the pathological tissue image of the analysis target region;
A clustering processing unit that performs a sinusoidal region extraction process by clustering on the pathological tissue image of the analysis target region that has been subjected to boundary enhancement processing by the boundary enhancement processing unit;
For the sinusoidal region extracted by the clustering processing unit, the computer is used as an sinusoidal region extraction processing unit including a false extraction region reduction processing step for performing erroneous extraction region reduction processing by a support vector machine (SVM). 38. The pathological image analysis program according to claim 37, wherein the pathological image analysis program is made to function. For the pathological tissue image of the analysis target region, a fiber probability calculation processing unit that performs fiber probability calculation processing by texture features;
With respect to the pathological tissue image of the analysis target region, a superpixel processing unit that performs superpixel processing for clustering into regions having similar colors,
A first stroma extraction processing unit that calculates an average fiber probability for each superpixel and extracts a superpixel having a high fiber probability as a stroma;
The computer is caused to function as a stromal region extraction processing unit comprising: a second stromal extraction processing unit that extracts superpixels having a large number of lymphocytes as a stroma. 1. A pathological image analysis program according to item 1.   In the quantification processing unit, the pathology tissue image of the parenchymal cell region extracted by the parenchymal cell region extraction processing unit is measured by measuring the width along the center line of the closed region figure of the parenchymal cell region, and quantified pathology 40. The pathological image analysis program according to any one of claims 37 to 39, wherein distance measurement processing that is output as a feature amount is performed. For the pathological tissue image of the parenchymal cell region extracted by the parenchymal cell region extraction processing unit, a thinning processing unit for thinning by obtaining the center line of the parenchymal cell region,
A cell line classification processing unit that classifies the parenchymal cell region in units of branches, with the branch from the branch to the branch of the thinned parenchymal cell region as a branch,
Distance measurement for measuring the distance from the center line of the parenchymal cell region to the boundary and outputting it as a quantified pathological feature amount for the pathological tissue image of the parenchymal parenchymal cell region classified by the cell cord classification processing unit 41. The pathological tissue image analysis program according to claim 40, wherein the computer is caused to function as a quantification processing unit including a processing unit.   42. The pathological image analysis program according to claim 41, wherein the quantification processing unit further calculates and outputs a nucleus density for each cell line or an NC ratio for each line as a pathological feature amount.