A method and system for pathological evaluation of artificial intelligence vascular irritation test

Through artificial intelligence deep learning technology, the identification and calculation of pathological images of vascular stimulation tests is solved, and the problems of time-consuming, labor-intensive and inconsistent human evaluation are achieved, fully quantitative pathological evaluation is improved, and the efficiency and accuracy of preclinical research are improved.

CN118692634BActive Publication Date: 2024-11-22ANLING BIOMEDICAL (SUZHOU) CO LTD +1
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411170558.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-22
Estimated Expiration
2044-08-26

AI Technical Summary

Technical Problem

In the prior art, artificial pathological evaluation is time-consuming, labor-intensive, inconsistent and semi-quantitative, and cannot accurately reflect changes in the degree of disease, resulting in prolonging preclinical research cycles and low reliability of results.

Method used

Using artificial intelligence methods, deep neural networks are used to identify and distinguish tissue structures and lesion areas in pathological images, subdividing vascular areas, identify vascular endothelial cells, and computing specific quantitative indicators to provide a fully quantitative pathological evaluation.

Benefits of technology

Achieve fully automatic pathological evaluation of intelligent vascular stimulation tests, improve the analysis speed and consistency of results, and generate multiple fully quantified indicators, which significantly improves the efficiency and accuracy of pathological diagnosis, and reduces labor costs and differences. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118692634B_ABST
    Figure CN118692634B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for pathological evaluation of artificial intelligence vascular irritation test, which relates to the field of drug clinical technology, including identifying and distinguishing tissue structures and lesion areas in pathological images; further subdividing the identified vascular area into vascular wall area and vascular cavity area; identifying vascular endothelial cells in the vascular wall; and calculating specific quantitative indicators based on the results of regional segmentation and cell nucleus recognition to evaluate the histopathological changes in the stimulated area. The method and system for pathological evaluation of artificial intelligence vascular irritation test has developed a fully automatic intelligent vascular irritation test pathological evaluation technology, which can significantly improve the speed of analysis, the results of the analysis are highly consistent, there is no difference in observation, and multiple fully quantitative indicators can be generated to better quantify the histological changes of experimental animals after contact with the test substance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of drug clinical technology, and in particular to a method and system for pathological evaluation of an artificial intelligence vascular irritation test. Background Art

[0002] Currently, in the preclinical evaluation of drugs, vascular irritation tests for intravenously infused drugs are usually combined with long-term toxicology tests, usually evaluating the limb veins of beagle dogs and crab-eating monkeys and the tail vein of SD rats. However, separate vascular irritation tests often use the marginal ear vein of rabbits for vascular irritation testing and evaluation.

[0003] For the vascular irritation test combined with the toxicology test, the scientific and animal welfare principles are followed. However, the writing of the vascular irritation report is the responsibility of the toxicologist, and the pathological changes of vascular irritation are diagnosed by the pathologist. The pathological evaluation in the vascular irritation test mainly reports: changes in the blood vessels at the administration site (vascular endothelium and vascular wall), changes around the blood vessels, and damage to the adjacent tissues. Usually, the pathological changes of the blood vessels involve: degeneration, necrosis, hyperplasia, and hypertrophy of the vascular endothelium; degeneration, necrosis, thickness, inflammation, etc. of each layer of the vascular wall; degeneration, necrosis, hemorrhage, inflammation, and fibrosis of the tissues around the blood vessels; degeneration, necrosis, hemorrhage, inflammation, and fibrosis of the adjacent tissues. For the above changes, pathologists can only make semi-quantitative diagnosis, not full quantification, such as the specific increase in vascular thickness, the specific number of endothelial cell hyperplasia, the bleeding area of ​​the blood vessels, and the extent of damage to the surrounding tissues. Different pathologists may have different evaluation results of vascular damage for the same drug.

[0004] In summary, the existing manual pathology evaluation has the following problems:

[0005] 1) Manual evaluation is time-consuming and labor-intensive, which prolongs the cycle of preclinical research;

[0006] 2) Manual evaluation results are inconsistent. There will be inconsistencies between evaluations by different pathologists or even by the same pathologist at different times.

[0007] 3) Manual evaluation uses a semi-quantitative grading system that can only obtain limited indicators, and each indicator can only be roughly divided into a limited number of levels, which cannot reflect the subtle changes in the severity of the disease.

[0008] Therefore, it is very necessary to develop artificial intelligence methods and systems for pathological evaluation of vascular irritation tests. Artificial intelligence can ensure the consistency of results, improve the speed of pathological diagnosis, accelerate the experimental process, reduce the requirements for pathology professionals, popularize the population and scope of pathological diagnosis, and reduce labor costs.

[0009] However, currently no method has been found to realize automated pathological evaluation of vascular irritation test using artificial intelligence image analysis technology. The present invention studies and improves the existing technology and its deficiencies, and provides a method and system for pathological evaluation of vascular irritation test using artificial intelligence. Summary of the invention

[0010] The purpose of the present invention is to provide a method and system for pathological evaluation of artificial intelligence vascular irritation test to solve the problems raised in the above-mentioned background technology.

[0011] To achieve the above object, the present invention provides the following technical solutions:

[0012] In a first aspect, a method for pathological evaluation of an artificial intelligence vascular irritation test is provided, comprising the following steps:

[0013] S1. Identify and differentiate tissue structures and lesion areas in pathological images;

[0014] S2, further subdividing the identified blood vessel area into a blood vessel wall area and a blood vessel lumen area;

[0015] S3, identify endothelial cells in the blood vessel wall;

[0016] S4. Based on the regional segmentation results and cell nucleus identification results, specific quantitative indicators are calculated to evaluate the histopathological changes in the stimulated area.

[0017] In a second aspect, a system for pathological evaluation of an artificial intelligence vascular irritation test is provided, which is applied to the method for pathological evaluation of an artificial intelligence vascular irritation test as described above, including:

[0018] Tissue structure and lesion area recognition module, vascular area fine recognition module, cell nucleus recognition module and lesion feature calculation module;

[0019] The tissue structure and lesion area recognition module is used to recognize and distinguish the tissue structure and lesion area in the pathological image, wherein the tissue structure includes the vascular area, the muscle area and other areas (mainly the dermis of the skin), and the lesion area includes the hemorrhage area and the inflammation area;

[0020] The blood vessel region fine identification module is used to further subdivide the identified blood vessel region into a blood vessel wall region and a blood vessel cavity region;

[0021] The cell nucleus recognition module is used to identify vascular endothelial cells in the blood vessel wall;

[0022] The lesion feature calculation module calculates specific quantitative indicators based on the regional segmentation results and cell nucleus recognition results obtained by the above three models, including:

[0023] Bleeding-related indicators: the proportion of bleeding area in the dermis, the maximum bleeding area in the dermis, the proportion of bleeding area in the vascular area, the maximum bleeding area in the vascular area, etc.

[0024] Inflammation-related indicators: the percentage of the inflammatory area in the dermis, the area of ​​the largest inflammatory area in the dermis, the percentage of the inflammatory area in the vascular area, the area of ​​the largest inflammatory area in the vascular area, etc.

[0025] Indicators related to vascular stenosis: ratio of vascular lumen to vascular area, average vascular wall thickness, width of the narrowest area of ​​vascular lumen, etc.

[0026] Vascular endothelial cell related indicators: vascular endothelial cell density, average size of vascular endothelial cells, etc.

[0027] Furthermore, the implementation scheme of the tissue structure and lesion area recognition module is: establishing a deep neural network 1 for segmenting skin tissue, and the segmented areas include bleeding areas, inflammation areas, blood vessel areas, muscle areas and other areas.

[0028] Furthermore, the segmentation model of the deep neural network 1 is trained based on the pathologist's delineation results of different tissue structure areas using the UperNet network structure (other structures, such as U-net, Deeplab, Mask-RCNN, etc., may also be used). The specific training process is as follows: a convolutional neural network with a Resnet50 structure is used as the encoder of the UperNet model, and the ImageNet pre-trained weights are loaded; the full-length digital pathology image is divided into 512*512 small blocks and input into the encoder to obtain multi-scale feature maps; then the decoder generates segmentation maps based on these feature maps, and completes the model training by minimizing different loss functions: Dice loss, weighted cross entropy, focal loss, etc.

[0029] Furthermore, the training of the deep neural network 1 is performed based on a resolution of 1 um / pixel, which can take into account both the cellular level information required for identifying the inflammatory area and the larger field of view required for identifying the complete blood vessel structure and muscle area.

[0030] Furthermore, the implementation scheme of the blood vessel region fine identification module is: establishing a deep neural network 2 for identifying the blood vessel wall and blood vessel cavity area.

[0031] Furthermore, the segmentation model of the deep neural network 2 is based on the pathologist's delineation results of different tissue structure areas, and also adopts the UperNet network structure. The specific training method is the same as the deep neural network 1. The difference is that the training is based on a resolution of 0.5um / pixel, which can identify the target area more finely.

[0032] Furthermore, the cell nucleus recognition module is implemented by establishing a deep neural network 3, segmenting all cell nuclei within a preset area at the junction of the vascular wall and the vascular cavity, and identifying the vascular endothelial cells therein based on the morphology of the cell nuclei.

[0033] Furthermore, the segmentation model of the deep neural network 3 is based on the pathologist's delineation results of the cell nucleus and is trained using the StarDist model. StarDist is a deep learning-based method designed specifically for object segmentation in biomedical images. The trained encoder is based on the U-Net network structure. During training, the distance transform is calculated based on the cell nucleus contour annotated by the doctor to generate a distance transform map, which encodes the distance from each pixel to the nearest cell nucleus boundary; the network learns to predict the segmentation result and the distance transform map; the distance transform loss helps to capture the spatial relationship and shape of the cell nucleus, and it penalizes the network according to the difference between the predicted distance map and the actual distance map, and finally outputs a probability map and a distance transform map, which are further processed to obtain the final segmentation result; after threshold segmentation of the predicted probability map and distance map, the contour of each cell nucleus is obtained;

[0034] Based on the fact that endothelial cells are mainly located in the inner wall of blood vessels, according to the fine segmentation results of the vascular wall and lumen, a 50-micron-wide area is set at the junction of the vascular wall and the lumen, and the cell nuclei are segmented only in this area;

[0035] As vascular endothelial cells are usually flat, the ratio of the major diameter to the minor diameter of each cell nucleus obtained by segmentation is measured. If the ratio is greater than 2, the cell nucleus is determined to be a vascular endothelial cell.

[0036] Furthermore, the lesion feature calculation module can choose to use any of the acquired measurement indicators alone to evaluate the histopathological changes in the stimulated area; it can also choose to build a model based on a set of training samples, through a series of measurement values ​​generated by the system and the results of manual scoring by doctors, and synthesize a series of measurement values ​​into a single measurement indicator to evaluate the histopathological changes in the stimulated area.

[0037] The present invention provides a method and system for pathological evaluation of artificial intelligence vascular irritation test, which has the following beneficial effects:

[0038] The present invention has developed a fully automatic intelligent vascular irritation test pathological evaluation technology, which can significantly improve the speed of analysis, with high consistency of analysis results, no observation differences, and can generate multiple fully quantitative indicators to better quantify the histological changes of experimental animals after exposure to the test substance. It has the following advantages:

[0039] 1) Apply artificial intelligence deep learning methods to the pathological evaluation of vascular irritation tests in preclinical experimental animals;

[0040] 2) Innovatively introduce a series of quantitative indicators to further refine the evaluation indicators and introduce more valuable information for preclinical research;

[0041] 3) Compared with the traditional manual interpretation method (time-consuming, labor-intensive, highly subjective, and poorly repeatable), this intelligent analysis system can provide more accurate, rapid, objective, and repeatable pathological evaluation results, which to a certain extent improves the work efficiency of pathologists and the implementation efficiency of preclinical research, and also reduces the impact of differences in manual evaluation by pathologists on research. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A logic block diagram of a system for pathological evaluation of an artificial intelligence vascular irritation test according to the present invention;

[0043] Figure 2 A schematic diagram of the correlation between the manual bleeding score and the artificial intelligence bleeding score of a system for pathological evaluation of an artificial intelligence vascular irritation test of the present invention;

[0044] Figure 3 This is a schematic diagram of the correlation between the manual inflammation score and the artificial intelligence inflammation score of a system for pathological evaluation of an artificial intelligence vascular irritation test of the present invention. DETAILED DESCRIPTION

[0045] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention. Example

[0046] A method for pathological evaluation of an artificial intelligence vascular irritation test, comprising the following steps:

[0047] S1. Identify and differentiate tissue structures and lesion areas in pathological images;

[0048] S2, further subdividing the identified blood vessel area into a blood vessel wall area and a blood vessel lumen area;

[0049] S3, identify endothelial cells in the blood vessel wall;

[0050] S4. Based on the regional segmentation results and cell nucleus identification results, specific quantitative indicators are calculated to evaluate the histopathological changes in the stimulated area. Example

[0051] like Figure 1-Figure 3As shown, a system for pathological evaluation of artificial intelligence vascular irritation test is applied to the method for pathological evaluation of artificial intelligence vascular irritation test as described above. The method steps of embodiment 1 can be executed by the system, and the system includes:

[0052] Tissue structure and lesion area recognition module, vascular area fine recognition module, cell nucleus recognition module and lesion feature calculation module;

[0053] The tissue structure and lesion area recognition module is used to identify and distinguish the tissue structure and lesion area in the pathological image, where the tissue structure includes the vascular area, muscle area and other areas (mainly the dermis of the skin), and the lesion area includes the hemorrhagic area and the inflammatory area;

[0054] The implementation scheme of the tissue structure and lesion area recognition module is as follows: a deep neural network 1 is established to segment skin tissue, and the segmented areas include bleeding areas, inflammation areas, blood vessel areas, muscle areas and other areas;

[0055] The segmentation model of deep neural network 1 is based on the pathologist's delineation results of different tissue structure areas and is trained using the UperNet network structure (other structures such as U-net, Deeplab, Mask-RCNN, etc. can also be used). The specific training process is as follows: a convolutional neural network with a Resnet50 structure is used as the encoder of the UperNet model, and the ImageNet pre-trained weights are loaded; the full digital pathology image is divided into 512*512 small blocks and input into the encoder to obtain multi-scale feature maps; then the decoder generates segmentation maps based on these feature maps, and completes the model training by minimizing different loss functions: Dice loss, weighted cross entropy, focal loss, etc.;

[0056] The training of the deep neural network 1 is based on a resolution of 1um / pixel, which can take into account both the cellular level information required to identify the inflammatory area and the larger field of view required to identify the intact vascular structure and muscle area;

[0057] The vascular region fine recognition module is used to further subdivide the recognized vascular region into a vascular wall region and a vascular cavity region;

[0058] The implementation scheme of the vascular region fine recognition module is as follows: establishing a deep neural network 2 for the recognition of the vascular wall and vascular cavity regions;

[0059] The segmentation model of deep neural network 2 is based on the delineation results of different tissue structure areas by pathologists. It also uses the UperNet network structure. The specific training method is the same as that of deep neural network 1. The difference is that the training is based on a resolution of 0.5um / pixel, which can identify the target area more precisely.

[0060] The cell nucleus recognition module is used to identify endothelial cells in the blood vessel wall;

[0061] The implementation scheme of the cell nucleus recognition module is as follows: a deep neural network 3 is established to segment all cell nuclei within a preset area at the junction of the vascular wall and the vascular cavity, and the vascular endothelial cells therein are identified based on the morphology of the cell nuclei;

[0062] The segmentation model of deep neural network 3 is based on the pathologist's delineation of the cell nucleus and is trained using the StarDist model. StarDist is a deep learning-based method designed for object segmentation in biomedical images. The trained encoder is based on the U-Net network structure. During training, the distance transform is calculated based on the cell nucleus outline annotated by the doctor to generate a distance transform map, which encodes the distance from each pixel to the nearest cell nucleus boundary; the network learns to predict the segmentation result and the distance transform map; the distance transform loss helps capture the spatial relationship and shape of the cell nucleus. It penalizes the network according to the difference between the predicted distance map and the actual distance map, and finally outputs a probability map and a distance transform map, which are further processed to obtain the final segmentation result; after threshold segmentation of the predicted probability map and distance map, the outline of each cell nucleus is obtained;

[0063] Based on the fact that endothelial cells are mainly located in the inner wall of blood vessels, according to the fine segmentation results of the vascular wall and lumen, a 50-micron-wide area is set at the junction of the vascular wall and the lumen, and the cell nuclei are segmented only in this area;

[0064] Since vascular endothelial cells are usually flat, the ratio of the long diameter to the short diameter of each cell nucleus obtained by segmentation is measured. If the ratio is greater than 2, the cell nucleus is determined to be a vascular endothelial cell.

[0065] The lesion feature calculation module calculates specific quantitative indicators based on the regional segmentation results and cell nucleus recognition results obtained by the above three models (i.e., deep neural network 1, 2, and 3), including:

[0066] Bleeding related indicators: the proportion of bleeding area in the dermis, the maximum bleeding area in the dermis, the proportion of bleeding area in the vascular area, the maximum bleeding area in the vascular area, etc. The correlation between the artificial intelligence bleeding score obtained in this embodiment and the traditional manual bleeding score is as follows: Figure 2 As shown;

[0067] Inflammation related indicators: the area ratio of the inflammation area in the dermis, the area of ​​the largest inflammation area in the dermis, the area ratio of the inflammation area in the vascular area, the area of ​​the largest inflammation area in the vascular area, etc.; the correlation between the artificial intelligence inflammation score obtained in this embodiment and the traditional manual inflammation score is as follows: Figure 3As shown;

[0068] Indicators related to vascular stenosis: ratio of vascular lumen to vascular area, average vascular wall thickness, width of the narrowest area of ​​vascular lumen, etc.

[0069] Vascular endothelial cell related indicators: vascular endothelial cell density, average size of vascular endothelial cells, etc.;

[0070] The lesion feature calculation module can choose to use any of the obtained measurement indicators alone to evaluate the histopathological changes in the stimulated area; it can also choose to build a model based on a set of training samples, through a series of measurement values ​​generated by this system and the results of manual scoring by doctors, and synthesize a series of measurement values ​​into a single measurement indicator to evaluate the histopathological changes in the stimulated area.

[0071] The embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

Claims

1. A method for pathological evaluation of artificial intelligence vascular irritation test, characterized in that: The following steps are involved: S1. Identify and differentiate tissue structures and lesion areas in pathological images; The tissue structure and lesion area recognition module is used to recognize and distinguish the tissue structure and lesion area in the pathological image, wherein the tissue structure includes the vascular area, the muscle area and other areas, and the lesion area includes the hemorrhagic area and the inflammatory area; The implementation scheme of the tissue structure and lesion area recognition module is as follows: a deep neural network 1 is established to segment skin tissue, and the segmented areas include bleeding areas, inflammation areas, blood vessel areas, muscle areas and other areas; The segmentation model of the deep neural network 1 is based on the delineation results of different tissue structure areas by pathologists and is trained using the UperNet network structure. The specific training process is as follows: a convolutional neural network with a Resnet50 structure is used as the encoder of the UperNet model, and the weights pre-trained by ImageNet are loaded; the full-length digital pathology image is divided into 512*512 small blocks and input into the encoder to obtain multi-scale feature maps; then the decoder generates segmentation maps based on these feature maps, and completes the model training by minimizing different loss functions; The training of the deep neural network 1 is based on a resolution of 1 um / pixel, which can take into account both the cellular level information required for identifying the inflammation area and the larger field of view required for identifying the complete blood vessel results and muscle areas; S2, further subdividing the identified blood vessel region into a blood vessel wall region and a blood vessel lumen region; the blood vessel region fine identification module is used to further subdivide the identified blood vessel region into a blood vessel wall region and a blood vessel lumen region; The implementation scheme of the blood vessel region fine identification module is as follows: establishing a deep neural network 2 for identifying the blood vessel wall and the blood vessel cavity region; The segmentation model of the deep neural network 2 is based on the delineation results of different tissue structure areas by pathologists, and also adopts the UperNet network structure. The specific training method is as follows: a convolutional neural network with a Resnet50 structure is used as the encoder of the UperNet model, and the weights pre-trained by ImageNet are loaded; the full-length digital pathology image is divided into 512*512 small blocks and input into the encoder to obtain multi-scale feature maps; then the decoder generates segmentation maps based on these feature maps, and completes the model training by minimizing different loss functions. The training of the deep neural network 2 is based on a resolution of 0.5um / pixel; S3, identifying vascular endothelial cells in the vascular wall; the cell nucleus recognition module is used to identify vascular endothelial cells in the vascular wall; The cell nucleus recognition module is implemented by: establishing a deep neural network 3, segmenting all cell nuclei within a preset area at the junction of the vascular wall and the vascular cavity, and identifying the vascular endothelial cells therein according to the morphology of the cell nuclei; The segmentation model of the deep neural network 3 is based on the pathologist's delineation of the cell nucleus and is trained using the StarDist model. The trained encoder is based on the U-Net network structure. During the training, the distance transformation is calculated based on the cell nucleus contour annotated by the doctor to generate a distance transformation map, which encodes the distance from each pixel to the nearest cell nucleus boundary; the network learns to predict the segmentation result and the distance transformation map; the distance transformation loss penalizes the network according to the difference between the predicted distance map and the actual distance map, and finally outputs a probability map and a distance transformation map, which are further processed to obtain the final segmentation result; after threshold segmentation of the predicted probability map and distance map, the contour of each cell nucleus is obtained; Based on the fact that endothelial cells are mainly located in the inner wall of blood vessels, according to the fine segmentation results of the vascular wall and lumen, a 50-micron-wide area is set at the junction of the vascular wall and the lumen, and the cell nuclei are segmented only in this area; By measuring the ratio of the long diameter to the short diameter of each cell nucleus obtained by segmentation, if the ratio is greater than 2, the cell nucleus is determined to be a vascular endothelial cell; S4. Calculate specific quantitative indicators based on the regional segmentation results and the cell nucleus recognition results to evaluate the histopathological changes in the stimulated area; the lesion feature calculation module calculates specific quantitative indicators based on the obtained regional segmentation results and the cell nucleus recognition results, including bleeding-related indicators, inflammation-related indicators, vascular stenosis-related indicators, and vascular endothelial cell-related indicators; The lesion feature calculation module can choose to use any of the obtained measurement indicators alone to evaluate the histopathological changes in the stimulated area; it can also choose to build a model based on a set of training samples, through a series of measurement values ​​and the results of manual scoring by doctors, and synthesize a series of measurement values ​​into a single measurement indicator to evaluate the histopathological changes in the stimulated area.

2. A system for pathological evaluation of artificial intelligence vascular irritation test, applied to the method for pathological evaluation of artificial intelligence vascular irritation test as claimed in claim 1, characterized in that: include: Tissue structure and lesion area recognition module, vascular area fine recognition module, cell nucleus recognition module and lesion feature calculation module.

Citation Information

Patent Citations

  • Intelligent analysis system and method for labial gland pathology

    CN113723441A

  • System for artificial intelligence assessment of non-alcoholic fatty liver animal model

    CN118098609A

  • Fundus pseudo focus detection method

    CN118115466A

  • Elderly patient ADR early warning model construction method, early warning system and storage medium

    CN118366683A