Intelligent pathology analysis platform based on reinforcement learning

Abstract

The invention relates to the field of smart medical treatment, and discloses a smart pathology analysis platform based on reinforcement learning, and the platform comprises an image segmentation unit which is used for determining an initial lesion region and a lesion type corresponding to the initial lesion region in a pathology image; the first adjusting unit is used for determining a boundary adjusting area according to the initial lesion area and the lesion type; wherein the boundary adjustment region is in the circumferential direction of the initial lesion region; and the second adjustment unit is used for adjusting the boundary of the boundary adjustment region according to the initial lesion region and the lesion type through a deep reinforcement learning model to obtain a target adjustment region, and taking a union set of the initial lesion region and the target adjustment region as a target lesion region. According to the method and the device, the boundary region of the lesion tissue in the pathological image is accurately adjusted, so that the effect of performing high-precision identification on the edge region of the lesion tissue is achieved.

Description

Technical Field

[0001] The present application relates to the field of smart medical technology, and more specifically, to a smart pathology analysis platform based on reinforcement learning. Background Art

[0002] Pathological images are pathological images taken by slicing tissues of the patient's lesion or by using medical imaging technology. Pathological images can reveal the cause and severity of the disease and are an important means of clinical diagnosis. Pathological images are an important tool used by doctors in disease diagnosis and research, and smart pathology is an emerging field that applies artificial intelligence technology to the field of pathology. One application of smart pathology is pathological image segmentation. Pathological image segmentation is the process of dividing pathological images into several independent non-intersecting regions based on certain similarity features of pathological images (such as brightness, color, texture, area, shape, position, local statistical features or spectral features, etc.). The relevant features show consistency or similarity in the same region, but show obvious differences between different regions, that is, there is a certain discontinuity in the pixels on the region boundary. Generally speaking, there is at least one region containing the target of interest in a meaningful image segmentation result. Reinforcement learning is a machine learning method that achieves a specific goal by maximizing the cumulative reward through interactive learning between the intelligent agent and the environment. In smart pathology, the introduction of reinforcement learning algorithms enables the pathological analysis system to automatically learn the decision-making strategy for judging whether there is cancer in the pathological image through interaction with the environment, thereby improving the accuracy and efficiency of pathological analysis. In the field of reinforcement learning, there are many mainstream algorithms that are suitable for different types of problems. Among them, deep reinforcement learning methods combine the advantages of deep learning and reinforcement learning, and can handle high-dimensional input data and complex decision-making tasks. Well-known deep reinforcement learning algorithms include Deep Q-Network (DQN), Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC). At present, in the field of pathological image analysis, there are few examples of applying reinforcement learning methods to analyze pathological images, and the effect of pathological image analysis based on reinforcement learning methods is poor.

[0003] Patent CN107665491B (application number: CN201710934902.6) provides a method and system for identifying pathological images, which can not only effectively retain low-dimensional and high-dimensional pathological features to identify lesion features, but also correspond lesion features to pixel positions of pathological images by fusing feature images output by at least two levels of feature processing units in a convolutional neural network and identifying pathological information in the pathological image based on the pathological features in the fused image. The edge area of ​​the lesion tissue in the pathological image has a certain significance for the pathological features, but the pathological image recognition method provided by patent CN107665491B cannot perform high-precision recognition of the edge area of ​​the lesion tissue in the pathological image, resulting in reduced use of the recognition results of the pathological image in pathological research. Summary of the invention

[0004] The purpose of this application is to provide an intelligent pathology analysis platform based on reinforcement learning, which solves the technical problem of not being able to perform high-precision identification of the edge area of ​​lesion tissue in pathology images, and achieves the technical effect of high-precision identification of the edge area of ​​lesion tissue in pathology images.

[0005] An embodiment of the present application provides an intelligent pathology analysis platform based on reinforcement learning, including: an image segmentation unit, used to determine an initial lesion area and a lesion type corresponding to the initial lesion area in a pathology image; a first adjustment unit, used to determine a boundary adjustment area according to the initial lesion area and the lesion type; wherein the boundary adjustment area is circumferentially of the initial lesion area; a second adjustment unit, used to adjust the boundary of the boundary adjustment area according to the initial lesion area and the lesion type through a deep reinforcement learning model to obtain a target adjustment area, and take the union of the initial lesion area and the target adjustment area as the target lesion area.

[0006] In one possible implementation, the second adjustment unit is also used to: segment the area close to the boundary of the initial lesion area according to the initial lesion area and the lesion type through a circumferential segmentation model of the lesion area, and obtain multiple circumferential areas within the area of ​​the boundary of the initial lesion area; wherein the circumferential area is used to indicate the boundary characteristics of the initial lesion area; determine the circumferential area center point of each circumferential area, and perform curve fitting on all circumferential area center points to obtain a fitted circumferential curve; wherein the fitted circumferential curve is a closed curve; determine a first boundary that encloses the initial lesion area and is parallel to the fitted circumferential curve, and use the area between the fitted circumferential curve and the first boundary as a boundary adjustment area; wherein the distance between the fitted circumferential curve and the first boundary is the first distance.

[0007] In another possible implementation, the second adjustment unit is also used to: determine the center point of the initial lesion area, and establish a two-dimensional coordinate system based on the center point of the initial lesion area; wherein the coordinate system divides the initial lesion area into four quadrant areas; determine the number of circumferential area center points in each quadrant area of ​​the coordinate system; when the number of circumferential area center points in at least one quadrant area in the coordinate system is less than a preset number, determine a second boundary that encloses the initial lesion area and is parallel to the boundary of the initial lesion area, and use the area enclosed by the initial lesion area and the second boundary as a boundary adjustment area; wherein the distance between the second boundary and the boundary of the initial lesion area is the second distance.

[0008] In another possible implementation, the second adjustment unit is further used to: determine the target adjustment area and the type of the target adjustment area according to the initial lesion area and the lesion type through an intelligent agent of a deep reinforcement learning model based on an action value strategy; wherein the value function of the deep reinforcement learning model is expressed by the following formula:

[0009] R = α × (T1 / T2)

[0010] Among them, R represents the reward value when the agent performs the boundary adjustment action, α represents the similarity between the type of the target adjustment area and the lesion type, T1 represents the feature value of the target adjustment area, and T2 represents the feature value of the initial lesion area.

[0011] In another possible implementation, the characteristic value of the target adjustment area includes a pixel average value of all pixels in the target adjustment area, and the characteristic value of the initial lesion area includes a pixel average value of all pixels in the initial lesion area.

[0012] In another possible implementation, the actions of the intelligent agent include: binarizing the initial lesion area; when the pixel value of the target pixel in the initial lesion area is greater than or equal to the preset pixel value, the pixel value of the target pixel is set to 1, and when the pixel value of the target pixel in the initial lesion area is less than the preset pixel value, the pixel value of the target pixel is set to 0; when the pixel values ​​of 26 adjacent pixels within 2 weeks around the target pixel are also 1, the target pixel and the 26 adjacent pixels around the target pixel are included in the target adjustment area.

[0013] In another possible implementation, the pathological image includes an intravascular OCT image, the initial lesion area includes a plaque area on the intravascular OCT image, and the circumferential area includes a suspected plaque erosion area in a circumferential direction of the plaque area.

[0014] In another possible implementation, when the pathological image is an intravascular OCT image, the value function of the deep reinforcement learning model is expressed by the following formula:

[0015] R = α × L (T1, T2)

[0016] Wherein, R represents the reward value when the agent performs the boundary adjustment action, α represents the similarity between the type of the target adjustment area and the lesion type, T1 represents the feature vector passing through the target adjustment area, T2 represents the feature vector of the initial lesion area, and L(T1, T2) represents the similarity between the feature vector T1 of the target adjustment area and the feature vector T2 of the initial lesion area; wherein, the feature vector T1 of the target adjustment area and the feature vector T2 of the initial lesion area both include the optical feature value, morphological feature value and three-dimensional continuity feature value of the tissue; wherein, the optical feature includes the optical intensity value and light attenuation value of the tissue, the optical intensity value includes the pixel value, contrast value and grayscale value in the image area, and the light attenuation value includes the change value of the pixel gradient in the image area; the morphological feature includes the difference value between the convex set of the image area and the image area, and the three-dimensional continuity feature value includes the optical features of the intravascular OCT image and the previous frame OCT image of the intravascular OCT image and the next frame OCT image of the intravascular OCT image, and the continuity feature value of the morphological feature.

[0017] In another possible implementation, the similarity between the type of the target adjustment area and the lesion type is determined by the following method: determine a lesion type vector composed of multiple lesion type probability values ​​corresponding to the initial lesion area; wherein the multiple lesion type probability values ​​corresponding to the initial lesion area are obtained through the output of the image segmentation unit; determine a circumferential area type vector composed of type probability values ​​of all circumferential areas corresponding to the target adjustment area; wherein the type probability value of each circumferential area is obtained through the output of the circumferential segmentation model of the lesion area, and the dimension of the circumferential area type vector and the lesion type vector are the same; determine the cosine similarity or Euclidean distance between the lesion type vector and the circumferential area type vector as the similarity between the type of the target adjustment area and the lesion type.

[0018] In another possible implementation manner, the value of the first distance is determined by the following method: the distance value between the center point of all the center points of the circumferential regions and the center point of the initial lesion region is taken as the value of the first distance.

[0019] In another possible implementation, the second distance is determined by the following formula: D2=β×Dmin, wherein D2 represents the value of the second distance, β represents the first lesion type coefficient determined according to the lesion type, and Dmin represents the value of the minimum distance from the boundary of the initial lesion area to the center point of the initial lesion area.

[0020] In another possible implementation, the second distance is determined by the following formula: D2=γ×Dt, wherein D2 represents the value of the second distance, γ represents the second lesion type coefficient determined according to the lesion type, and Dt represents the distance value between the center point of all circumferential area center points and the center point of the initial lesion area.

[0021] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0022] The embodiment of the present application provides an intelligent pathology analysis platform based on reinforcement learning, including: an image segmentation unit, used to determine the initial lesion area and the lesion type corresponding to the initial lesion area in the pathology image; a first adjustment unit, used to determine the boundary adjustment area according to the initial lesion area and the lesion type; wherein the boundary adjustment area is in the circumferential direction of the initial lesion area; a second adjustment unit, used to adjust the boundary of the boundary adjustment area according to the initial lesion area and the lesion type through a deep reinforcement learning model to obtain a target adjustment area, and use the union of the initial lesion area and the target adjustment area as the target lesion area. The embodiment of the present application improves the recognition accuracy of the image in the edge area of ​​the lesion area by effectively adjusting the edge area of ​​the initial lesion area in the pathology image that is initially identified, improves the recognition accuracy of the lesion area of ​​the pathology image, and improves the effect of the pathology image in pathology research. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0024] Figure 1 A schematic diagram of the system structure of an intelligent pathology analysis platform in an embodiment of the present application;

[0025] Figure 2 A schematic diagram of a process of identifying a pathological image by an intelligent pathological analysis platform in an embodiment of the present application;

[0026] Figure 3 A schematic diagram of a process of identifying a pathological image by another intelligent pathological analysis platform in an embodiment of the present application;

[0027] Figure 4 A schematic diagram of a process of identifying a pathological image by another intelligent pathological analysis platform in an embodiment of the present application;

[0028] In the figure, 100, intelligent pathology analysis platform; 110, image segmentation unit; 120, first adjustment unit; 130, second adjustment unit; 200, pathology image; 210, initial lesion area; 211, circumferential area; 211a, center point of circumferential area; 212, fitting circumferential curve; 213, first boundary; 214, second boundary; 220, boundary adjustment area; 230, target adjustment area; 240, target lesion area. DETAILED DESCRIPTION

[0029] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0030] It should be noted that when a component or structure is referred to as being "fixed to" or "disposed on" another component or structure, it may be directly on the other component or structure or indirectly on the other component or structure. When a component or structure is referred to as being "connected to" another component or structure, it may be directly connected to the other component or structure or indirectly connected to the other component or structure.

[0031] It should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or a component or structure referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0032] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0033] Patent CN107665491B (application number: CN210710934902.6) provides a method and system for identifying pathological images, which can not only effectively retain low-dimensional and high-dimensional pathological features to identify lesion features, but also correspond lesion features to pixel positions of pathological images by fusing feature images output by at least two levels of feature processing units in a convolutional neural network and identifying pathological information in the pathological image based on the pathological features in the fused image. The edge area of ​​the lesion tissue in the pathological image has a certain significance for the pathological features, but the pathological image recognition method provided by patent CN107665491B cannot perform high-precision recognition of the edge area of ​​the lesion tissue in the pathological image, resulting in a reduced use effect of the recognition result of the pathological image in pathological research.

[0034] Based on the above reasons, the embodiment of the present application provides an intelligent pathology analysis platform based on reinforcement learning, including: an image segmentation unit, used to determine the initial lesion area and the lesion type corresponding to the initial lesion area in the pathology image; a first adjustment unit, used to determine the boundary adjustment area according to the initial lesion area and the lesion type; wherein the boundary adjustment area is in the circumferential direction of the initial lesion area; a second adjustment unit, used to adjust the boundary of the boundary adjustment area according to the initial lesion area and the lesion type through a deep reinforcement learning model to obtain a target adjustment area, and take the union of the initial lesion area and the target adjustment area as the target lesion area. The embodiment of the present application improves the recognition accuracy of the image within the edge area of ​​the lesion area by effectively adjusting the edge area of ​​the initial lesion area in the pathology image that is initially identified, improves the recognition accuracy of the lesion area of ​​the pathology image, and improves the effect of the pathology image in pathology research.

[0035] In some scenarios, an intelligent pathology analysis platform based on reinforcement learning in an embodiment of the present application can be applied to the image recognition of pathological images taken by slicing tissues at the patient's lesion site, which can improve the pathological research effect of tissue slices at the patient's lesion site.

[0036] In other scenarios, an intelligent pathology analysis platform based on reinforcement learning in an embodiment of the present application can also be applied to pathological images captured using medical imaging technology, for example, it can be applied to pathological research in OCT images, which can improve the effect of pathological research on medical images.

[0037] The following is a detailed description of an intelligent pathology analysis platform based on reinforcement learning provided in an embodiment of the present application with reference to specific examples.

[0038] Figure 1 Schematic diagram of the system structure of a smart pathology analysis platform in an embodiment of the present application. Figure 1As shown, the intelligent pathology analysis platform 100 in the embodiment of the present application includes an image segmentation unit 110, a first adjustment unit 120 and a second adjustment unit 130. The image segmentation unit 110, the first adjustment unit 120 and the second adjustment unit 130 are described in detail below.

[0039] like Figure 1 As shown, the embodiment of the present application includes an image segmentation unit 110, which is used to determine an initial lesion area and a lesion type corresponding to the initial lesion area in a pathological image when in operation.

[0040] Figure 2 FIG. 1 is a schematic diagram of a process of identifying a pathological image by an intelligent pathological analysis platform in an embodiment of the present application. Figure 2 As shown in FIG. a, when the pathological image 200 is segmented by the image segmentation unit 110, an initial lesion region 210 and a lesion type corresponding to the initial lesion region 210 can be determined in the pathological image 200.

[0041] Exemplarily, the lesion type corresponding to the initial lesion region 210 may be marked by outputting a label of the initial lesion region 210 by the image segmentation unit 110 .

[0042] Exemplarily, the lesion type may be an early lesion, a mid-stage lesion, or a late-stage lesion. The lesion type may also be a benign lesion or a malignant lesion. The embodiment of the present application does not limit the type of pathological image and the lesion type corresponding to the pathological image.

[0043] For example, the pathological image may be a pathological image obtained by slicing tissue of a patient's diseased area, or may be a pathological image obtained by using medical imaging technology (eg, an OCT image).

[0044] like Figure 1 As shown, the embodiment of the present application further includes a first adjustment unit 120, which is used to determine a boundary adjustment region according to the initial lesion region and the lesion type, wherein the boundary adjustment region is in the circumferential direction of the initial lesion region.

[0045] like Figure 2 As shown in FIG. b, when the first adjustment unit 120 is working, it is used to determine the boundary adjustment area 220 according to the initial lesion area 210 and the lesion type. The boundary adjustment area 220 is circumferentially of the initial lesion area 210, that is, the boundary adjustment area 220 includes the area circumferentially of the initial lesion area 210.

[0046] Pathological image segmentation is the process of dividing a pathological image into several independent non-overlapping regions based on certain similarity features of the pathological image (such as brightness, color, texture, area, shape, position, local statistical features or spectral features, etc.). The areas near the boundaries of these independent regions are of certain value to the analysis of pathological images and the study of pathology. Therefore, it is of great significance to perform high-precision identification of the areas near the boundaries of independent regions in pathological images.

[0047] The embodiment of the present application determines the boundary adjustment area 220, and then identifies the circumferential area of ​​the initial lesion area 210 based on the boundary adjustment area 220, so as to more accurately identify the area near the boundary of the initial lesion area 210, thereby improving the analysis effect of the pathological image.

[0048] like Figure 1 As shown, the embodiment of the present application also includes a second adjustment unit 130. When the second adjustment unit 130 is working, it is used to adjust the boundary of the boundary adjustment area according to the initial lesion area and the lesion type through a deep reinforcement learning model to obtain a target adjustment area, and take the union of the initial lesion area and the target adjustment area as the target lesion area.

[0049] After the first adjustment unit 120 identifies the boundary adjustment area 220, the second adjustment unit 130 can analyze the pathological image in the boundary adjustment area 220, thereby improving the processing effect of the pathological information in the boundary adjustment area 220 and improving the effect of revealing the pathological mechanism in the pathological image.

[0050] At work, if Figure 2 As shown in the process from Figure b to Figure c, the second adjustment unit 130 adjusts the boundary of the boundary adjustment area 220 according to the initial lesion area 210 and the lesion type through a deep reinforcement learning model to obtain a target adjustment area 230, and takes the union of the initial lesion area 210 and the target adjustment area 230 as the target lesion area 240.

[0051] When the pathological image 200 is identified, after the target lesion area 240 is obtained, a more accurate study can be conducted based on the pathological mechanism of the target lesion area 240 , and a more accurate pathological study can also be conducted on the edge area of ​​the lesion tissue based on the more accurate target lesion area 240 .

[0052] The beneficial effect brought about by the above-mentioned implementation method is that when the intelligent pathology analysis platform in the embodiment of the present application is used, the pathology image is segmented for the first time through the image segmentation unit, and then the initial lesion area and lesion type are determined. The first segmentation of the image segmentation unit realizes the coarse segmentation of the pathology image, thereby improving the initial recognition efficiency of the image.

[0053] The beneficial effect brought about by the above-mentioned implementation method is that the boundary adjustment area that needs to be adjusted is determined by the first adjustment unit, and then the boundary adjustment area is adjusted for the second time by the second adjustment unit to obtain the target lesion area, and then the union of the target lesion area and the initial lesion area is taken as the target lesion area. By identifying the boundary adjustment area, the recognition effect of the pathological image is improved according to the boundary adjustment area that needs to be adjusted, which can improve the accuracy of the subsequent application of the recognition results of the pathological image for pathological analysis.

[0054] The beneficial effect brought about by the above-mentioned implementation method is that after achieving efficient positioning of the boundary adjustment area, the boundary adjustment area is finely adjusted through the deep reinforcement learning model, thereby improving the accuracy of identifying the local area, and the effect of adjusting the local area is improved through the deep reinforcement learning model, and finally an accurate recognition result of the boundary area of ​​the pathological image is obtained.

[0055] In some implementations, the second adjustment unit 120 is further configured to execute S310 to S330 , which will be described in detail below.

[0056] S310: Segment the region near the boundary of the initial lesion region 210 according to the initial lesion region 210 and the lesion type using the lesion region circumferential segmentation model to obtain a plurality of circumferential regions 211 within the region of the boundary of the initial lesion region 210. The circumferential regions 211 are used to indicate boundary features of the initial lesion region 210.

[0057] Figure 3 FIG. 1 is a schematic diagram of another process of identifying a pathological image by an intelligent pathological analysis platform in an embodiment of the present application. Figure 3 As shown in Figure b11, the circumferential region 211 may be a closed region identified based on the features of the image. After the circumferential region 211 is identified, the boundary features of the initial lesion region 210 can be indicated based on the circumferential region 211, thereby improving the effect of the subsequent pathological analysis based on the circumferential region of the initial lesion region 210.

[0058] For example, the plurality of circumferential regions 211 within the boundary of the initial lesion region 210 may be regions of interest that play a greater role in the analysis of pathological mechanisms in the analysis of pathological images. For example, the circumferential region 211 may be a region where cancerous cells are relatively concentrated.

[0059] Exemplarily, the circumferential segmentation model of the lesion area may be a trained convolutional neural network model, and the circumferential segmentation model of the lesion area may be capable of segmenting the area close to the boundary of the initial lesion area 210 according to the initial lesion area 210 and the lesion type.

[0060] Exemplarily, when training a circumferential segmentation model for a lesion area, the training set and the test set may be pathological images marked with the lesion area and the circumferential area around the lesion area, the lesion area having a corresponding lesion type label, and the circumferential area having a corresponding lesion type label.

[0061] S320, determining the circumferential region center point of each circumferential region, and performing curve fitting on all circumferential region center points to obtain a fitted circumferential curve, wherein the fitted circumferential curve is a closed curve.

[0062] like Figure 3 As shown in FIG. b12 , after the circumferential regions 211 are identified, the circumferential region center point 211 a of each circumferential region 211 can be determined, and then the distribution of the circumferential regions 211 can be determined based on the distribution of the circumferential region center points 211 a.

[0063] Exemplarily, the circumferential region center point 211a may be a geometric center point, or the circumferential region center point 211a may be a pixel weighted center point determined according to a pixel value.

[0064] like Figure 3 As shown in FIG. b12, after obtaining the circumferential region center point 211a, a fitting circumferential curve 212 can be obtained by curve fitting all circumferential region center points 211a, and then the distribution of the circumferential region 211 of interest can be simulated by the fitting circumferential curve 212. The fitting circumferential curve 212 is a closed curve.

[0065] After the circumferential region 211 is identified, the boundary features of the initial lesion region 210 can be indicated according to the circumferential region 211, thereby improving the effect of subsequent pathological analysis based on the circumferential region of the initial lesion region 210.

[0066] S330, determining a first boundary that surrounds the initial lesion area and is parallel to the fitting circumferential curve, and using the area between the fitting circumferential curve and the first boundary as the boundary adjustment area, wherein the distance between the fitting circumferential curve and the first boundary is the first distance.

[0067] like Figure 3 As shown in FIG. b12 , after the fitting circumferential curve 212 is obtained, a first boundary 213 that surrounds the initial lesion area 210 and is parallel to the fitting circumferential curve 212 can be determined.

[0068] like Figure 3 As shown in FIG. b12 , the distance between the fitting circumferential curve 212 and the first boundary 213 is the first distance D1 .

[0069] After obtaining the first boundary 213, the area between the fitted circumferential curve 212 and the first boundary 213 can be used as the boundary adjustment area 220. The boundary adjustment area 220 is the area of ​​interest. The boundary adjustment area 220 can be used as the circumferential characteristic area of ​​the initial lesion area 210 in the pathological image 200, thereby characterizing the image pathological characteristics of the initial lesion area 210 in the circumferential direction.

[0070] Exemplarily, when fitting the circumferential curve 212 , the curve fitting may be performed by an interpolation method to obtain the fitted circumferential curve 212 .

[0071] The beneficial effect brought about by the above-mentioned implementation method is that the circumferential area of ​​the initial lesion area is identified through the circumferential segmentation model of the lesion area, and then the pathological characteristics of the circumferential area of ​​the initial lesion area can be obtained based on the circumferential area, which can improve the recognition effect of the pathological image and facilitate subsequent pathological analysis.

[0072] The beneficial effect brought about by the above-mentioned implementation method is that the circumferential area of ​​the initial lesion area is identified through the circumferential segmentation model of the lesion area, and the fitting circumferential curve and the first boundary are determined, and then the boundary adjustment area is determined through the center point of the circumferential area, thereby realizing the utilization of the distribution characteristics of the circumferential area around the initial lesion area, improving the scientific nature of determining the boundary adjustment area, being able to improve the recognition accuracy of the boundary adjustment area, being able to improve the effect of subsequent adjustments to the boundary adjustment area, and being able to improve the analysis and recognition effect of the pathological image as a whole.

[0073] In some implementations, the second adjustment unit 130 is further configured to execute S410 to S430, which will be described below.

[0074] S410: Determine the center point of the initial lesion area, and establish a two-dimensional coordinate system based on the center point of the initial lesion area, wherein the coordinate system divides the initial lesion area into four quadrants.

[0075] Figure 4 FIG. 1 is a schematic diagram of another process of identifying a pathological image by an intelligent pathological analysis platform in an embodiment of the present application. Figure 4 As shown in FIG. b21 , the center point o of the initial lesion region 210 can be determined, and then a two-dimensional xoy coordinate system can be established based on the center point o of the initial lesion region 210 .

[0076] Exemplarily, when determining the center point o of the initial lesion region 210 , the geometric center point of the initial lesion region 210 may be determined as the center point o, or the center point o of the initial lesion region 210 may be calculated by weighting according to pixel values.

[0077] like Figure 4As shown in Figure b21, after the two-dimensional Xoy coordinate system is established, the two-dimensional Xoy coordinate system divides the initial lesion area into four quadrant areas: quadrant area 1, quadrant area 2, quadrant area 3 and quadrant area 4.

[0078] S420: Determine the number of circumferential area center points in each quadrant area of ​​the coordinate system.

[0079] In quadrant area 1, quadrant area 2, quadrant area 3 and quadrant area 4, the number of circumferential area center points 211a in each quadrant area can be determined respectively, and then the distribution of the circumferential area 211 in each quadrant area can be determined according to the number of circumferential area center points 211a in each quadrant area.

[0080] S430: When the number of circumferential area center points in at least one quadrant area in the coordinate system is less than a preset number, determine a second boundary that surrounds the initial lesion area and is parallel to the boundary of the initial lesion area, and use the area surrounded by the initial lesion area and the second boundary as the boundary adjustment area. The distance between the second boundary and the boundary of the initial lesion area is the second distance.

[0081] When the number of circumferential regions 211 in a quadrant is too small, when the boundary adjustment region is determined through the process from S310 to S330, the fitting circumferential curve 212 may be distorted or fail to fit due to the small number of circumferential regions 211.

[0082] For example, Figure 4 As shown in Figure b22, when the preset number is 3, when the number of circumferential area center points 211a in each quadrant area in the xoy coordinate system is less than the preset number of 3, when the boundary adjustment area is determined through the process in S310 to S330, the fitting circumferential curve 212 may be distorted because the number of circumferential areas 211 is too small.

[0083] like Figure 4 As shown in FIG. b22 , at this time, by determining a second boundary 214 that encloses the initial lesion area 210 and is parallel to the boundary of the initial lesion area 210 , the area enclosed by the initial lesion area 210 and the second boundary 214 can be used as the boundary adjustment area 220 .

[0084] like Figure 4 As shown in FIG. b22 , the distance between the second boundary 214 and the boundary of the initial lesion area 210 is the second distance D2 .

[0085] Exemplarily, the preset number may be 3, 6 or 10.

[0086] The beneficial effect brought about by the above-mentioned implementation method is that if the number of circumferential areas is small, the boundary of the boundary adjustment area is directly determined by the boundary of the initial lesion area, thereby avoiding the distortion caused by determining the boundary adjustment area by the circumferential area when the number of circumferential areas is too small, and improving the robustness of pathological image recognition.

[0087] The beneficial effect brought about by the above-mentioned implementation method is that by dividing the initial lesion area into four quadrant areas, the number of circumferential area center points in each quadrant area is counted. If the number of circumferential area center points in a quadrant area is too small, the boundary of the initial lesion area is used to determine the boundary of the boundary adjustment area, thereby improving the efficiency of the statistical effect of the distribution of the circumferential area center points and improving the image processing efficiency.

[0088] In some implementations, the second adjustment unit 130 is further configured to execute step S510, which will be described in detail below.

[0089] S510. The intelligent agent of the deep reinforcement learning model determines the target adjustment area and the type of the target adjustment area based on the action value strategy and the initial lesion area and the lesion type.

[0090] Since the deep reinforcement learning method combines the advantages of deep learning and reinforcement learning, it can handle high-dimensional input data and complex decision-making tasks. When working, the deep reinforcement learning model can improve the effect of image recognition. Among them, the deep reinforcement learning model executes actions and updates states based on action value strategies.

[0091] When the deep reinforcement learning model is working, it can determine the target adjustment region 230 and the type of the target adjustment region 230 according to the initial lesion region 210 and the lesion type. The initial lesion region 210 and the lesion type have been obtained in the output of the above-mentioned image segmentation unit 110.

[0092] like Figure 2 As shown in Figure c, through the deep reinforcement learning model, the initial lesion area 210 and the lesion type, the boundary of the boundary adjustment area 220 is adjusted to obtain the target adjustment area 230, and the union of the initial lesion area 210 and the target adjustment area 230 is used as the target lesion area 240.

[0093] For example, the value function of the deep reinforcement learning model can be expressed by the following formula:

[0094] R=α×T1 / T2

[0095] Among them, R represents the reward value when the agent performs the boundary adjustment action, α represents the similarity between the type of the target adjustment area 230 and the lesion type, T1 represents the feature value of the target adjustment area 230, and T2 represents the feature value of the initial lesion area 210.

[0096] The beneficial effect brought about by the above-mentioned implementation method is that, through the deep reinforcement learning intelligent agent, it is possible to combine the advantages of the deep reinforcement learning model, accurately obtain the adjusted target adjustment area and the type of the target adjustment area, and improve the effect of identifying the target adjustment area and the type of the target adjustment area.

[0097] The beneficial effect brought about by the above implementation method is that the action of the intelligent agent is evaluated through a reward function based on the similarity of the type and the characteristic value of the region, thereby improving the recognition effect of the target adjustment region.

[0098] In some implementations, the characteristic value of the target adjustment area includes a pixel average value of all pixels in the target adjustment area, and the characteristic value of the initial lesion area includes a pixel average value of all pixels in the initial lesion area.

[0099] For example, the pixel average value of all pixels in the target adjustment area 230 may be calculated and used as the feature value of the target adjustment area.

[0100] Exemplarily, the pixel average value of all pixels in the initial lesion area may also be calculated, and the pixel average value of all pixels in the initial lesion area may be used as the characteristic value of the initial lesion area.

[0101] The beneficial effect brought about by the above-mentioned implementation method is that by taking the pixel average value of all pixels in the initial lesion area or the pixel average value of all pixels in the target adjustment area as the characteristic value of the corresponding area, efficient calculation of the characteristics of the corresponding area can be achieved, thereby improving the calculation efficiency of the target adjustment area.

[0102] In some implementations, the actions of the above-mentioned intelligent agent include S610 and S620, and S610 and S620 are described in detail below.

[0103] S610, binarizing the initial lesion area, wherein when the pixel value of the target pixel in the initial lesion area is greater than or equal to the preset pixel value, the pixel value of the target pixel is set to 1, and when the pixel value of the target pixel in the initial lesion area is less than the preset pixel value, the pixel value of the target pixel is set to 0.

[0104] By performing binarization processing on the initial lesion region 210 , the initial lesion region 210 can be converted into a grayscale image, and thus the initial lesion region 210 can be processed efficiently.

[0105] Exemplarily, the pixel value of the target pixel in the initial lesion area 210 is recorded in the form of RGB pixel value, and the RGB pixel value can be converted into pixel value by the formula Gray=0.299R+0.587G+0.114B, where 0.299, 0.587 and 0.114 are the weights of the pixel values ​​of the three channels.

[0106] Exemplarily, when the preset pixel value is 0.8, when the pixel value of the target pixel in the initial lesion area 210 is greater than or equal to the preset pixel value 0.8, the pixel value of the target pixel is set to 1; when the pixel value of the target pixel in the initial lesion area is less than the preset pixel value 0.8, the pixel value of the target pixel is set to 0.

[0107] S620: When the pixel values ​​of 26 adjacent pixels within 2 weeks around the target pixel are all 1, the target pixel and the 26 adjacent pixels around the target pixel are included in the target adjustment area.

[0108] When the intelligent agent is working, it can first determine the target pixel point, and then detect the pixel values ​​of the 26 adjacent pixel points within 2 weeks around the target pixel point. When the pixel values ​​of the 26 adjacent pixel points within 2 weeks around the target pixel point are also 1, the target pixel point and the 26 adjacent pixel points around the target pixel point are included in the target adjustment area 230. Subsequently, the action of the intelligent agent can be evaluated according to the value function to determine the next action of the intelligent agent.

[0109] The beneficial effect brought about by the above-mentioned implementation method is that the initial lesion area 210 is firstly processed quickly according to the binarization process, so as to be used as input in the subsequent determination of the action of the intelligent agent, thereby realizing efficient adjustment of the target adjustment area through the deep reinforcement learning model.

[0110] In some implementations, the pathological image includes an intravascular OCT image, the initial lesion area includes a plaque area on the intravascular OCT image, and the circumferential area includes a suspected plaque erosion area in the circumferential direction of the plaque area.

[0111] Atherosclerosis is an inflammatory fibroproliferative disease. Plaques formed on the endothelium of blood vessels can cause vascular stenosis or thrombosis, leading to ischemia and even severe acute coronary syndromes (ACS). In the diagnosis of intravascular plaques, the plaque erosion area is an important cause of plaque instability. Plaque erosion often manifests itself in the form of intact plaque fibrous cap, irregular surface, and superficial calcification. Therefore, it is of great significance to identify the plaque erosion area.

[0112] In the application scenario where the pathological image to be processed is an intravascular OCT image, the plaque area on the intravascular OCT image can be first identified in the intravascular OCT image, and the plaque area is used as the initial lesion area. In subsequent processing, the suspected plaque erosion area in the circumferential direction of the plaque area is identified as a circumferential area, and the circumferential area is adjusted by the second adjustment unit 130 to achieve accurate identification of the suspected plaque erosion area.

[0113] The beneficial effect brought about by the above-mentioned implementation method is that in the process of recognizing intravascular OCT images, when identifying the initial lesion area formed by the plaque area, the suspected plaque erosion area can be regarded as a circumferential area and efficiently identified, thereby achieving a more accurate recognition effect on acute coronary syndrome and improving the processing effect of pathological images of acute coronary syndrome.

[0114] In some implementations, when the pathological image is an intravascular OCT image, the value function of the deep reinforcement learning model is expressed by the following formula:

[0115] R = α × L (T1, T2)

[0116] Among them, R represents the reward value when the agent performs a boundary adjustment action, α represents the similarity between the type of the target adjustment area and the lesion type, T1 represents the feature vector passing through the target adjustment area, T2 represents the feature vector of the initial lesion area, and L(T1, T2) represents the similarity between the feature vector T1 of the target adjustment area and the feature vector T2 of the initial lesion area.

[0117] During use, when processing intravascular OCT images, in order to improve the processing effect of the intravascular OCT images, the value function may be optimized.

[0118] When in use, α represents the similarity between the type of the target adjustment area 230 and the lesion type. The action value of the intelligent agent is evaluated by the similarity between the type of the target adjustment area 230 and the lesion type. The action of the intelligent agent can be effectively guided according to the adjustment target of the target adjustment area, so as to select the target adjustment area 230 that is similar to the type of the initial lesion area 210, thereby achieving the purpose of accurately determining the target adjustment area 230.

[0119] When in use, by determining the feature vector T1 representing the target adjustment area 230, and determining the feature vector T2 representing the initial lesion area 210, and by calculating the similarity L(T1, T2) between the feature vector T1 representing the target adjustment area 230 and the feature vector T2 representing the initial lesion area 210, an effective comparison of the features of the target adjustment area 230 and the initial lesion area 210 is achieved, thereby achieving the purpose of accurately determining the target adjustment area 230.

[0120] In some implementations, the feature vector T1 of the target adjustment region 230 and the feature vector T2 of the initial lesion region 210 both include optical feature values, morphological feature values, and three-dimensional continuity feature values ​​of the tissue.

[0121] The optical features may include optical intensity values ​​and light attenuation values ​​of the tissue, the optical intensity values ​​include pixel values, contrast values ​​and grayscale values ​​within the image area, and the light attenuation values ​​include pixel gradient change values ​​within the image area.

[0122] The morphological feature may include a convex set of the image region and a difference value between the image region.

[0123] The three-dimensional continuity feature value may include continuity feature values ​​of the optical features and morphological features of the intravascular OCT image and the previous frame OCT image and the next frame OCT image of the intravascular OCT image.

[0124] The beneficial effect brought about by the above-mentioned implementation method is that by calculating the similarity between the type of the target adjustment area 230 and the lesion type, as well as the vector similarity calculated by calculating the feature vector of the target adjustment area 230 and the feature vector of the initial lesion area 210, and combining the similarity between the type of the target adjustment area 230 and the lesion type and the similarity between the feature vector T1 of the target adjustment area and the feature vector T2 of the initial lesion area to calculate the reward value of the value function, a scientific adjustment of the action of the intelligent agent is achieved.

[0125] The beneficial effect brought about by the above implementation method is that the feature vector of the target adjustment area 230 and the feature vector of the initial lesion area 210 are calculated by using the parameters of multiple blood vessel images in the intravascular OCT image, thereby improving the accuracy of identifying suspected plaque erosion areas in the intravascular OCT image.

[0126] In some implementations, the method for determining the similarity between the type of the target adjustment region and the lesion type includes S710 to S730, and S710 to S730 are described in detail below.

[0127] S710, determining a lesion type vector composed of a plurality of lesion type probability values ​​corresponding to the initial lesion region, wherein the plurality of lesion type probability values ​​corresponding to the initial lesion region are obtained through output of an image segmentation unit.

[0128] When working, when the image segmentation unit 110 outputs the initial lesion region 210 , it can also output multiple lesion type probability values ​​corresponding to the initial lesion region 210 .

[0129] For example, the multiple lesion type probability values ​​may be that the probability that the initial lesion region 210 is type A lesion is 0.81, the probability that the initial lesion region 210 is type B lesion is 0.10, and the probability that the initial lesion region 210 is type C lesion is 0.09.

[0130] When in use, a lesion type vector composed of multiple lesion type probability values ​​corresponding to the initial lesion area 210 can be determined, and the lesion type vector can be (0.81, 010, 0.09).

[0131] S720: Determine a circumferential region type vector composed of type probability values ​​of all circumferential regions corresponding to the target adjustment region, wherein the type probability value of each circumferential region is obtained through the output of the lesion region circumferential segmentation model.

[0132] When in use, by determining the circumferential region type vector composed of the type probability values ​​of all circumferential regions 211 corresponding to the target adjustment region 230, subsequent calculations can be performed based on the circumferential region type vector reflecting the type probability values ​​of all circumferential regions 211, and by calculating the similarity between the circumferential region type vector and the lesion type vector, the similarity between the initial lesion region 210 and the target adjustment region 230 can be calculated.

[0133] The type probability value of each circumferential region 211 is obtained through the output of the circumferential segmentation model of the lesion region.

[0134] Exemplarily, when the number of circumferential regions 211 is 3, the probability of the first circumferential region is 0.79, the probability of the second circumferential region is 0.15, the probability of the third circumferential region is 0.06, and the circumferential region type vector may be (0.79, 0.15, 0.06).

[0135] S730: Determine the cosine similarity or Euclidean distance between the lesion type vector and the circumferential region type vector as the similarity between the type of the target adjustment region and the lesion type.

[0136] When in use, the cosine similarity or Euclidean distance between the lesion type vector and the circumferential area type vector is determined, and then the cosine similarity or Euclidean distance is used as the similarity between the type of the target adjustment area and the lesion type, thereby realizing the quantitative calculation of the similarity between the lesion type vector and the circumferential area type vector.

[0137] The beneficial effect brought about by the above-mentioned implementation method is that, through vector calculation, the similarity between the type of the target adjustment area and the lesion type is determined, thereby achieving reasonable guidance of the action of the intelligent body, and then all circumferential areas 211 that are most similar to the type of the initial lesion area 210 can be determined. The target adjustment area 230 is determined by all circumferential areas 211, so the target adjustment area 230 that is most similar to the type of the initial lesion area 210 can be determined, thereby achieving efficient guidance of the action of the intelligent body.

[0138] In some implementations, the value of the first distance is determined by the following method: the distance value between the center point of all circumferential area center points 211a and the center point of the initial lesion area 210 is used as the value of the first distance D1.

[0139] The beneficial effect brought about by the above-mentioned implementation method is that the distance value between the center point of the circumferential area center point 211a and the center point of the initial lesion area 210 is used as the value of the first distance D1, and then the value of the first distance D1 can be calculated by the distribution of multiple circumferential areas 211, so that the value of the first distance D1 can be scientifically determined.

[0140] The beneficial effect brought about by the above-mentioned implementation method is that if the distribution of multiple circumferential regions 211 around the initial lesion region 210 is more uniform, it means that the positions of the multiple circumferential regions 211 of the target adjustment region 230 are more accurately determined, and the value of the first distance D1 is smaller at this time; if the distribution of multiple circumferential regions 211 around the initial lesion region 210 is more uneven, it means that the positions of the multiple circumferential regions 211 of the target adjustment region 230 are more inaccurate, and the value of the first distance D1 is larger at this time, thereby realizing the scientific calculation of the value of the first distance D1 through the distribution of multiple circumferential regions 211, thereby improving the recognition effect of the target adjustment region 230.

[0141] In some implementations, the second distance D2 is determined by the following formula:

[0142] D2=β×Dmin

[0143] Wherein, D2 represents the value of the second distance, β represents the first lesion type coefficient determined according to the lesion type, and Dmin represents the value of the minimum distance from the boundary of the initial lesion area to the center point of the initial lesion area.

[0144] Exemplarily, the first lesion type coefficient β determined according to the lesion type can be set to different values ​​according to different lesion types, and then the first lesion type coefficient β can be determined according to different lesion types, thereby achieving the purpose of calculating the second distance D2 according to different lesion types.

[0145] Exemplarily, the first lesion type coefficient β may be 0.3, 0.6 or 0.8.

[0146] When in use, by calculating the value Dmin of the minimum distance from the boundary of the initial lesion area to the center point of the initial lesion area, the second distance D2 can be determined according to the boundary of the initial lesion area.

[0147] The beneficial effect brought about by the above-mentioned implementation method is that the second distance is determined by the lesion type coefficient and the minimum distance from the boundary of the initial lesion area to the center point of the initial lesion area, and then the second distance can be calculated by the lesion type coefficient corresponding to different lesion types and the minimum distance from the boundary of the initial lesion area to the center point of the initial lesion area. The second distance can be determined according to the lesion type and the size of the lesion area, which realizes the reasonable calculation of the width of the target boundary area and improves the recognition effect of the target adjustment area.

[0148] In some implementations, the second distance is determined by the following formula:

[0149] D2=γ×Dt

[0150] Wherein, D2 represents the value of the second distance, γ represents the second lesion type coefficient determined according to the lesion type, and Dt represents the distance value between the center point of all circumferential area center points and the center point of the initial lesion area.

[0151] When in use, the second lesion type coefficient γ is determined according to the lesion type, and the second distance D2 can be determined by the second lesion type coefficient γ.

[0152] When in use, the second distance D2 is determined by the distance value Dt between the center point of all circumferential area center points 211 a and the center point of the initial lesion area 210 .

[0153] It should be noted that the first lesion type coefficient β and the second lesion type coefficient γ corresponding to the lesion type of the same initial lesion region 210 are different.

[0154] The beneficial effect brought about by the above-mentioned implementation method is that the second distance is determined by the second lesion type coefficient determined by the lesion type and the distance between the center point of all circumferential areas and the center point of the initial lesion area, and then the second distance can be calculated according to the lesion type and the distribution of all circumferential areas, thereby realizing the reasonable calculation of the width of the target boundary area and improving the recognition effect of the target adjustment area.

[0155] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A smart pathology analysis platform based on reinforcement learning, characterized in that: include: An image segmentation unit, used for determining an initial lesion area and a lesion type corresponding to the initial lesion area in a pathological image; A first adjustment unit, configured to determine a boundary adjustment region according to the initial lesion region and the lesion type; wherein the boundary adjustment region is in a circumferential direction of the initial lesion region; The second adjustment unit is used to adjust the boundary of the boundary adjustment area according to the initial lesion area and the lesion type through a deep reinforcement learning model to obtain a target adjustment area, and take the union of the initial lesion area and the target adjustment area as the target lesion area.

2. The intelligent pathology analysis platform based on reinforcement learning as claimed in claim 1, characterized in that: The second adjustment unit is further used for: By using a circumferential segmentation model of the lesion region, segmenting the region close to the boundary of the initial lesion region according to the initial lesion region and the lesion type, to obtain a plurality of circumferential regions within the region of the boundary of the initial lesion region; wherein the circumferential regions are used to indicate boundary features of the initial lesion region; Determine the circumferential region center point of each circumferential region, and perform curve fitting on all circumferential region center points to obtain a fitted circumferential curve; wherein the fitted circumferential curve is a closed curve; A first boundary is determined which surrounds the initial lesion area and is parallel to the fitted circumferential curve, and an area between the fitted circumferential curve and the first boundary is used as the boundary adjustment area; wherein the distance between the fitted circumferential curve and the first boundary is a first distance.

3. The intelligent pathology analysis platform based on reinforcement learning as claimed in claim 2, characterized in that: The second adjustment unit is further used for: Determine the center point of the initial lesion area, and establish a two-dimensional coordinate system according to the center point of the initial lesion area; wherein the coordinate system divides the initial lesion area into four quadrant areas; Determining the number of the circumferential region center points in each quadrant region of the coordinate system; When the number of center points of the circumferential area in at least one quadrant area in the coordinate system is less than a preset number, a second boundary is determined that encloses the initial lesion area and is parallel to the boundary of the initial lesion area, and the area enclosed by the initial lesion area and the second boundary is used as the boundary adjustment area; wherein the distance between the second boundary and the boundary of the initial lesion area is the second distance.

4. The intelligent pathology analysis platform based on reinforcement learning as claimed in claim 3, characterized in that: The second adjustment unit is further used for: The target adjustment area and the type of the target adjustment area are determined according to the initial lesion area and the lesion type by an intelligent agent of a deep reinforcement learning model based on an action value strategy; wherein the value function of the deep reinforcement learning model is expressed by the following formula: R = α × (T1 / T2) Among them, R represents the reward value when the agent performs the boundary adjustment action, α represents the similarity between the type of the target adjustment area and the lesion type, T1 represents the feature value of the target adjustment area, and T2 represents the feature value of the initial lesion area.

5. The intelligent pathology analysis platform based on reinforcement learning as claimed in claim 4, characterized in that: The characteristic value of the target adjustment area includes a pixel average value of all pixels in the target adjustment area, and the characteristic value of the initial lesion area includes a pixel average value of all pixels in the initial lesion area.

6. The intelligent pathology analysis platform based on reinforcement learning as claimed in claim 5, characterized in that: The actions of the agent include: Binarization is performed on the initial lesion area; when the pixel value of the target pixel point in the initial lesion area is greater than or equal to a preset pixel value, the pixel value of the target pixel point is set to 1, and when the pixel value of the target pixel point in the initial lesion area is less than the preset pixel value, the pixel value of the target pixel point is set to 0; When the pixel values ​​of 26 adjacent pixels within 2 weeks around the target pixel are all 1, the target pixel and the 26 adjacent pixels around the target pixel are included in the target adjustment area.

7. The intelligent pathology analysis platform based on reinforcement learning as claimed in claim 3, characterized in that: The pathological image includes an intravascular OCT image, the initial lesion area includes a plaque area on the intravascular OCT image, and the circumferential area includes a suspected plaque erosion area in a circumferential direction of the plaque area.

8. The intelligent pathology analysis platform based on reinforcement learning as claimed in claim 7, characterized in that: When the pathological image is an intravascular OCT image, the value function of the deep reinforcement learning model is expressed by the following formula: R = α × L (T1, T2) Wherein, R represents the reward value when the agent performs the boundary adjustment action, α represents the similarity between the type of the target adjustment area and the lesion type, T1 represents the feature vector passing through the target adjustment area, T2 represents the feature vector of the initial lesion area, and L(T1, T2) represents the similarity between the feature vector T1 of the target adjustment area and the feature vector T2 of the initial lesion area; Among them, the feature vector T1 of the target adjustment area and the feature vector T2 of the initial lesion area both include optical feature values, morphological feature values ​​and three-dimensional continuity feature values ​​of the tissue; wherein the optical feature includes the optical intensity value and light attenuation value of the tissue, the optical intensity value includes the pixel value, contrast value and grayscale value within the image area, and the light attenuation value includes the change value of the pixel gradient within the image area; the morphological feature includes the difference value between the convex set of the image area and the image area, and the three-dimensional continuity feature value includes the continuity feature values ​​of the optical features and morphological features of the intravascular OCT image and the previous frame OCT image of the intravascular OCT image and the next frame OCT image of the intravascular OCT image.

9. The intelligent pathology analysis platform based on reinforcement learning as claimed in claim 8, characterized in that: The similarity between the type of the target adjustment area and the type of the lesion is determined by the following method: Determine a lesion type vector composed of a plurality of lesion type probability values ​​corresponding to the initial lesion area; wherein the plurality of lesion type probability values ​​corresponding to the initial lesion area are obtained through the output of the image segmentation unit; Determine a circumferential region type vector composed of type probability values ​​of all circumferential regions corresponding to the target adjustment region; wherein the type probability value of each circumferential region is obtained by outputting the circumferential segmentation model of the lesion region, and the circumferential region type vector has the same dimension as the lesion type vector; The cosine similarity or Euclidean distance between the lesion type vector and the circumferential region type vector is determined as the similarity between the type of the target adjustment region and the lesion type.

10. The intelligent pathology analysis platform based on reinforcement learning according to any one of claims 3 to 9, characterized in that: The value of the first distance is determined by the following method: The distance value between the center point of all the circumferential area center points and the center point of the initial lesion area is used as the value of the first distance.

Images