Tumor pathological image segmentation method based on U-Net neural network

Abstract

The invention is suitable for the field of medical image analysis, and provides a tumor pathological image segmentation method based on a U-Net neural network, and the method comprises the steps: extracting the gray, texture and shape multi-dimensional features of tumor cells through a U-Net encoder, constructing a feature cluster, building an invasion direction probability model based on the cluster, analyzing the morphological features of a necrotic region, and speculating the growth speed. And dynamically adjusting the segmentation sensitivity threshold of the decoder, and finally outputting a three-channel segmentation map containing a tumor core region, a false envelope invasion region and a capillary invasion region. According to the scheme, through multi-dimensional feature fusion and biological behavior modeling, precise characterization of tumor heterogeneity is achieved, the adaptability of a segmentation model to different invasion active areas is improved through a dynamic sensitivity regulation mechanism, a segmentation result with anatomical positioning and biological evaluation values is provided for clinic, and the segmentation accuracy is improved. And diagnosis and treatment decision of tumors are effectively assisted. The method is easy and convenient to operate and high in automation degree and has remarkable clinical application value.

Description

Technical Field

[0001] The present invention belongs to the field of medical image analysis, and in particular relates to a tumor pathology image segmentation method based on a U-Net neural network. Background Art

[0002] In the field of medical image analysis, accurate segmentation of tumor pathology images is the key basis for clinical diagnosis and treatment plan formulation. Especially for malignant tumors such as liver cancer and bladder cancer, the boundary definition of the tumor area in the pathology image, the assessment of the degree of invasion and the analysis of the growth status directly affect the patient's prognosis. Currently, image segmentation technology based on deep learning (such as U-Net neural network) has been widely used in pathology image analysis. Its end-to-end feature learning capability has significantly improved the segmentation efficiency and automation level compared with traditional manual segmentation or manual feature-based methods. With the explosive growth of medical imaging data and the development needs of precision medicine, how to deeply integrate the biological behavior characteristics of tumors (such as invasion path, growth rate) with image segmentation technology has become an important research direction to improve the accuracy of pathological diagnosis.

[0003] Existing technologies mainly implement tumor pathology image segmentation based on the U-Net neural network architecture. It obtains multi-scale features through an encoder-decoder structure and jump connections to complete pixel-level segmentation of the tumor area. Some solutions attempt to optimize the segmentation results by combining texture features or morphological analysis, but they usually only stay at a single-scale feature extraction, fail to fully integrate the grayscale, texture, shape and other multi-dimensional features of tumor cells, and lack quantitative analysis of biological characteristics such as tumor invasion direction and necrotic area morphology. In addition, the segmentation sensitivity threshold of the decoder in the existing technology is mostly fixed, which makes it difficult to adapt to the characteristics of strong heterogeneity of tumor tissue and significant differences in biological behavior in different regions. Summary of the Invention

[0004] The purpose of the present invention is to provide a tumor pathology image segmentation method based on a U-Net neural network, aiming to solve the technical problems existing in the prior art identified in the background technology.

[0005] The present invention is implemented as follows: a tumor pathology image segmentation method based on a U-Net neural network, the method comprising:

[0006] The pathological image is input into the U-Net neural network for encoding, and the grayscale features, texture features, and shape features of the tumor cells in the image are extracted at different levels of the encoder. The hierarchical features are then integrated to construct a multidimensional feature cluster that characterizes the distribution and morphology of cancer cells.

[0007] Based on the multidimensional feature cluster, a probability model of tumor cell invasion direction is established to quantify the probability of tumor cells breaking through the pseudocapsule and invading the liver parenchyma, and to simulate the potential path of tumor cell spread;

[0008] In the multidimensional feature cluster, the size, density and edge definition morphological features of the necrotic area in the pathological image are identified and analyzed, and the relative growth rate of the tumor at different stages is estimated based on the morphological features of the necrotic area;

[0009] Dynamically adjusting the sensitivity threshold of the U-Net decoder segmentation based on the invasion probability information, relative growth rate, and tumor region boundary regularity information output by the invasion direction probability model;

[0010] The input pathological image is segmented using the U-Net decoder with dynamic sensitivity adjustment, and the segmentation result map is output.

[0011] As a further embodiment of the present invention, the construction of a multi-dimensional feature cluster characterizing the distribution and morphology of cancer cells specifically includes:

[0012] Using the convolutional and pooling layers of the U-Net encoder, the basic visual features of the image, including pixel grayscale value distribution and local contrast information, are extracted in the first 1-2 downsampling levels of the encoder to form the basic visual feature layer;

[0013] At the 3rd and 4th downsampling levels of the encoder, the contrast of the gray-level co-occurrence matrix and the local binary pattern histogram are calculated to extract the intermediate texture feature layer describing the density and structural heterogeneity of tumor cells.

[0014] At encoder level 5 and above, high-level morphological feature layers of tumor cell clusters are extracted through semantic segmentation mask generation and contour fitting;

[0015] Through jump connections, the basic visual feature layer, the intermediate texture feature layer and the high-level morphological feature layer are channel-joined and fused with 1×1 convolution to generate a multidimensional feature cluster containing multi-scale spatial information and semantic information.

[0016] As a further embodiment of the present invention, the quantification of the probability of tumor cells breaking through the pseudocapsule and invading the liver parenchyma and simulating the potential path of tumor cell spread specifically includes:

[0017] Define the pseudocapsule area feature template and the perivascular area feature template;

[0018] Calculating the matching degree between the multidimensional feature cluster and the pseudocapsule region feature template to generate a probability map of tumor cells breaking through the pseudocapsule;

[0019] Calculating the matching degree between the multi-dimensional feature cluster and the feature template of the perivascular region, and combining the vascular direction information to generate a path probability map of tumor cells spreading along the venous wall;

[0020] fuse the breakthrough pseudo-envelope probability map and the path probability map to form a comprehensive tumor invasion direction probability model output.

[0021] As a further scheme of the present application, a matching degree is calculated to generate a breakthrough pseudo-envelope probability map of tumor cells, specifically:

[0022] A sliding window is traversed to calculate the matching degree of the multi-dimensional feature cluster at each position with the pseudo-envelope region feature template:

[0023]

[0024] wherein MatchScore(x,y) represents the matching degree of the multi-dimensional feature cluster at position (x,y) with the pseudo-envelope region feature template, F cluster (x,y) represents the multi-dimensional feature cluster vector at position (x,y), and T capsule represents the pseudo-envelope region feature template.

[0025] The matching degree is normalized:

[0026]

[0027] wherein min(Score) represents the minimum value of the matching degree of the multi-dimensional feature cluster with the pseudo-envelope region feature template, and max(Score) represents the maximum value of the matching degree of the multi-dimensional feature cluster with the pseudo-envelope region feature template.

[0028] Output P break (x,y) as the breakthrough pseudo-envelope probability map.

[0029] As a further scheme of the present application, the matching degree is calculated and combined with the blood vessel direction information to generate a path probability map of tumor cells along the vein wall, specifically comprising:

[0030]

[0031] wherein VascMatch(x,y) is the matching degree of the multi-dimensional feature cluster at position (x,y) with the blood vessel surrounding region feature template, T vessel is the blood vessel surrounding region feature template, and σ is the Gaussian kernel width.

[0032] A blood vessel skeleton line direction field θ(x,y) is extracted to generate a direction weight map W direction (x,y):

[0033] W direction (x,y)=cos 2 (φ(x,y)-θ(x,y));

[0034] Where φ(x,y) represents the link direction from position (x,y) to the centerline of the tumor;

[0035] Fusion generates path probability graph P spread (x,y):

[0036] P spread (x,y)=VascMatch(x,y)×W direction (x,y).

[0037] As a further embodiment of the present invention, the inferring the relative growth rate of the tumor at different stages based on the morphological characteristics of the necrotic area specifically includes:

[0038] Detecting and segmenting necrotic regions in the multidimensional feature cluster;

[0039] Calculating the area ratio of the necrotic area and the standard deviation of the internal pixel intensity to quantify the size and density of the necrotic area;

[0040] Calculating the gradient intensity distribution of the boundary of the necrotic area to evaluate the edge clarity of the necrotic area;

[0041] Based on the tumor growth dynamics model, the necrotic area is classified into stages in combination with the size, density and edge clarity of the necrotic area, and the relative growth rate estimation of the tumor area is output.

[0042] As a further solution of the present invention, the dynamic adjustment of the sensitivity threshold of the U-Net decoder segmentation specifically includes:

[0043] Comprehensively analyze the path probability map, growth rate estimation results, and edge clarity, and set corresponding recognition threshold intervals respectively.

[0044] Based on the threshold interval, high-probability areas, fast-growing areas, and areas with fuzzy boundaries are read and marked as high-sensitivity demand areas; low-probability areas, slow-growing areas, and areas with regular boundaries are read and marked as low-sensitivity demand areas;

[0045] Matching activation function thresholds are set for the high-sensitivity requirement area and the low-sensitivity requirement area respectively.

[0046] As a further solution of the present invention, the segmentation of the input pathological image using the U-Net decoder with dynamic sensitivity adjustment specifically includes:

[0047] Bilinear interpolation upsampling is used for high-sensitivity areas, and Sigmoid activation function is applied to enhance weak signals;

[0048] For low-sensitivity areas, the nearest neighbor upsampling is used, and the Sigmoid activation function is applied to suppress noise;

[0049] The basic visual feature layer and intermediate texture feature layer extracted by the encoder are weightedly spliced ​​with the dynamic sensitivity feature map of the corresponding layer of the decoder through jump connections, where the dynamic feature fusion weight coefficient of the weighted splicing is W fuse :

[0050] W fuse =α·P break +β·GrowthSpeed+γ·(1-GradSharpness);

[0051] Among them, α, β, γ are weight coefficients, P break is the probability of breaking through the pseudocapsule, GrowthSpeed ​​is the growth speed estimation result, and GradSharpness is a quantitative indicator of the edge clarity of the necrotic area;

[0052] Output a three-channel segmentation map, which includes: tumor core area, pseudocapsule invasion area and microvascular invasion area.

[0053] The beneficial effects of the present invention are:

[0054] This solution leverages the multi-level feature extraction mechanism of a U-Net neural network to construct a multidimensional feature cluster encompassing grayscale, texture, and shape. Combined with a probabilistic model of tumor invasion direction and analysis of necrotic growth rates, it enables quantitative characterization of tumor biological behavior in liver and bladder cancer pathology images. A mechanism for dynamically adjusting the decoder sensitivity threshold enhances segmentation accuracy in areas with high invasion probability, rapid growth, and blurred boundaries, while suppressing noise interference in low-risk areas, significantly improving the segmentation model's adaptability to tumor heterogeneity.

[0055] The final output of the three-channel segmentation map of the tumor core area, pseudocapsule invasion area and microvascular invasion area can not only accurately define the anatomical boundaries of the tumor, but also reflect the invasive activity and growth stage of tumor cells. It provides multi-dimensional information with both spatial positioning and biological significance for the clinical pathological grading, surgical margin assessment and metastasis risk prediction of liver cancer and bladder cancer, promotes the upgrade of pathological image segmentation from simple morphological description to biological function analysis, and provides innovative technical support for precision medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flowchart of a tumor pathology image segmentation method based on a U-Net neural network provided in an embodiment of the present invention;

[0057] Figure 2 A flowchart of constructing a multidimensional feature cluster that characterizes the distribution and morphology of cancer cells provided in an embodiment of the present invention;

[0058] Figure 3A flowchart for quantifying the probability of tumor cells breaking through the pseudocapsule and invading the liver parenchyma, and simulating potential pathways for tumor cell spread, provided by an embodiment of the present invention;

[0059] Figure 4 A flow chart of estimating the relative growth rate of a tumor at different stages based on the morphological characteristics of the necrotic area provided in an embodiment of the present invention;

[0060] Figure 5 A flowchart for dynamically adjusting the sensitivity threshold of U-Net decoder segmentation provided by an embodiment of the present invention;

[0061] Figure 6 A flowchart of segmenting an input pathological image using a U-Net decoder with dynamic sensitivity adjustment is provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0063] Figure 1 The flowchart of the tumor pathology image segmentation method based on the U-Net neural network provided in the embodiment of the present invention is as follows: Figure 1 As shown, the method includes:

[0064] S100: Input the pathological image into a U-Net neural network for encoding, extract the grayscale features, texture features, and shape features of tumor cells in the image at different levels of the encoder, and fuse the hierarchical features to construct a multidimensional feature cluster representing the distribution and morphology of cancer cells.

[0065] In the first 1-2 downsampling levels, convolution operations are used to extract basic visual features such as pixel grayscale value distribution and local contrast. These features serve as the cornerstone, laying the underlying visual foundation for subsequent more advanced feature extraction, enabling the model to perceive the most basic light and dark changes and local differences in the image.

[0066] At the 3rd and 4th downsampling levels, the contrast is calculated with the help of the gray-level co-occurrence matrix, and the intermediate texture features describing the density and structural heterogeneity of tumor cells are extracted through local binary pattern histogram analysis. This can capture the unique texture patterns of tumor cells in their arrangement and structure, providing key information for identifying the tissue characteristics of the tumor.

[0067] At the 5th level and above, semantic segmentation mask generation and contour fitting are used to extract the high-level morphological feature layer of tumor cell clusters, forming a complete morphological cognition, enabling the model to understand the overall shape and contour characteristics of tumor cell clusters.

[0068] The base visual feature layer, the intermediate texture feature layer and the high-level morphological feature layer are spliced and fused by 1x1 convolution through skip connection. This fusion method can generate a multi-dimensional feature cluster containing multi-scale spatial information and semantic information, so that the model can grasp the subtle local features of the tumor and understand its overall morphology and spatial distribution.

[0069] The base visual features used in this step ensure the model's perception of the basic elements of the image. The intermediate texture features delve into the organizational characteristics of tumor cells, and the high-level morphological features understand the structure of the tumor as a whole. The fusion of the three enables the model to recognize the tumor from different dimensions.

[0070] In liver tumor pathological image segmentation, this mechanism can accurately identify the gray difference between tumor cells and normal liver tissue, capture the texture features of abnormal dense arrangement of tumor cells, and outline the irregular morphology of tumor masses, thereby providing comprehensive and accurate feature support for subsequent establishment of invasion direction probability model and speculation of growth rate.

[0071] This multi-level feature representation greatly improves the model's ability to express the complex morphology and heterogeneity of tumors, making the segmentation results more accurately reflect the actual situation of the tumor and provide more reliable basis for clinical diagnosis. In actual application, for different types of tumor pathological images, such as liver cancer, bladder tumor, etc. Pathological sections, this step can accurately capture the unique features of each type of tumor through multi-level feature extraction, laying a solid foundation for subsequent accurate segmentation and analysis, and showing strong adaptability and effectiveness.

[0072] As shown in Figure 2 , the multi-dimensional feature cluster representing the distribution and morphology of cancer cells includes:

[0073] S110, using the convolutional layers and pooling layers of the U-Net encoder, extract the base visual features of the image at the first 1-2 down-sampling levels of the encoder, including pixel gray value distribution and local contrast information, to form the base visual feature layer;

[0074] S120, at the third to fourth down-sampling levels of the encoder, extract the intermediate texture feature layer describing the arrangement density and structural heterogeneity of tumor cells by calculating the contrast of the gray level co-occurrence matrix and the local binary pattern histogram;

[0075] S130, at the fifth level and above of the encoder, extract the high-level morphological feature layer of tumor cell clusters through semantic segmentation mask generation and contour fitting;

[0076] S140, through skip connections, channel splicing and 1×1 convolution fusion are performed on the basic visual feature layer, the intermediate texture feature layer, and the high-level morphological feature layer to generate a multidimensional feature cluster containing multi-scale spatial information and semantic information.

[0077] S200, establishing a probability model of tumor cell invasion direction based on the multidimensional feature cluster, quantifying the probability of tumor cells breaking through the pseudocapsule and invading the liver parenchyma, and simulating potential paths of tumor cell spread;

[0078] In this step, when establishing a probability model of tumor cell invasion direction based on multidimensional feature clusters, the feature templates of the pseudocapsule area and the perivascular area are first defined to provide a priori structural reference for the quantitative analysis of tumor invasion behavior.

[0079] These feature templates can accurately match the invasion-related feature patterns in pathological images. When calculating the matching degree between the multidimensional feature cluster and the pseudocapsule template, a sliding window is used to traverse each position in the image, and the feature similarity is quantified using vector dot product and norm normalization. The generated pseudocapsule breakthrough probability map can intuitively display the potential risk areas for tumor cells to break through the capsule.

[0080] The matching degree calculation with the template around the blood vessels combines the Gaussian kernel function with the blood vessel direction information. It not only measures the feature similarity through the exponential decay function, but also incorporates the spatial correlation between the tumor diffusion direction and blood vessel distribution into the analysis through the directional weight map. The resulting path probability map can effectively simulate the potential trajectory of tumor diffusion along the venous wall.

[0081] The fusion output of these two probability maps enables the model to construct a comprehensive probability model of tumor invasion direction based on the two key dimensions of capsule breakthrough and vascular invasion, achieving multi-dimensional dynamic prediction of tumor invasion behavior. The significant advantage of this step lies in the precise quantification and spatial prediction of tumor invasion behavior achieved through feature template matching and multi-source probability fusion mechanisms.

[0082] The pseudocapsule breakthrough probability map can capture the key sites of tumor infiltration into the surrounding liver parenchyma, while the vascular diffusion path probability map can reveal the potential channels for tumor metastasis through blood circulation. The combination of the two enables the model to simulate the "dual path" of tumor invasion.

[0083] In liver tumor pathology image analysis, this mechanism can accurately identify the risk of breakthrough in weak pseudocapsule areas and simultaneously track the possible pathways of tumor cell spread along the portal vein or hepatic vein, providing a quantitative basis for clinical assessment of tumor metastasis risk. This modeling approach, which combines structural features (pseudocapsule) with biological behavioral features (vascular invasion), overcomes the limitations of traditional segmentation methods that focus solely on regional boundaries, enabling segmentation results to more deeply reflect the biological behavior of tumors.

[0084] In practical applications, for tumor types with different invasive characteristics (such as capsular invasion of hepatocellular carcinoma and vascular infiltration of bladder tumors), this model can accurately predict their invasion direction and diffusion path through adaptive matching of feature templates, providing key imaging support for the formulation of personalized treatment plans, and demonstrating unique clinical application value in tumor pathology analysis.

[0085] like Figure 3 As shown, the method quantifies the probability of tumor cells breaking through the pseudocapsule and invading the liver parenchyma, and simulates the potential path of tumor cell spread, specifically including:

[0086] S210, defining a pseudocapsule region feature template and a perivascular region feature template;

[0087] S220, calculating a matching degree between the multidimensional feature cluster and the pseudocapsule region feature template to generate a probability map of tumor cells breaking through the pseudocapsule;

[0088] S230, calculating a matching degree between the multi-dimensional feature cluster and the feature template of the perivascular region, and generating a path probability map of tumor cells spreading along the venous wall in combination with the vascular direction information;

[0089] S240, fusing the pseudocapsule breakthrough probability map and the path probability map to form a comprehensive tumor invasion direction probability model output.

[0090] In this step, the matching degree is calculated to generate a probability map of tumor cells breaking through the pseudocapsule, specifically:

[0091] The sliding window is traversed to calculate the matching degree between the multidimensional feature cluster at each position and the feature template of the pseudo-encapsulation area:

[0092]

[0093] Among them, MatchScore(x,y) represents the matching degree between the multidimensional feature cluster at position (x,y) and the feature template of the pseudocapsule area, F cluster (x,y) represents the feature cluster vector at position (x,y), T capsule Represents the characteristic template of the pseudocapsule area;

[0094] Normalize the matching degree:

[0095]

[0096] Among them, min(Score) represents the minimum value of the matching degree between the multidimensional feature cluster and the pseudocapsule region feature template, and max(Score) represents the maximum value of the matching degree between the multidimensional feature cluster and the pseudocapsule region feature template;

[0097] Output P break (x,y) is used as the probability map of breaking through the false envelope.

[0098] In this step, the matching degree calculation is performed and combined with the blood vessel direction information to generate a path probability map of tumor cells spreading along the vein wall, which specifically includes:

[0099]

[0100] Among them, VascMatch(x,y) is the matching degree between the multidimensional feature cluster at position (x,y) and the feature template of the perivascular area, T vessel is the feature template of the perivascular area, σ is the Gaussian kernel width;

[0101] Extract the vascular skeleton line direction field θ(x,y) and generate the direction weight map W direction (x,y):

[0102] W direction (x,y)=cos 2 (φ(x,y)-θ(x,y));

[0103] Where φ(x,y) represents the link direction from position (x,y) to the centerline of the tumor;

[0104] Fusion generates path probability graph P spread (x,y):

[0105] P spread (x,y)=VascMatch(x,y)×W direction (x,y).

[0106] S300, identifying and analyzing the size, density, and edge definition morphological features of necrotic areas in the pathological image in the multidimensional feature cluster, and estimating the relative growth rates of the tumor at different stages based on the morphological features of the necrotic areas;

[0107] In this step, when identifying the morphological features of the necrotic area based on multidimensional feature clustering, the necrotic area is first accurately located from the multidimensional features through feature clustering and threshold segmentation technology. This process can make the necrotic area stand out from normal tissues and tumor cells.

[0108] Subsequently, by calculating the area ratio and pixel intensity standard deviation of the necrotic area, a quantitative analysis of the size and density of the necrotic area is achieved - the area ratio reflects the spatial distribution scale of the necrotic area, and the pixel intensity standard deviation reveals the degree of heterogeneity of its internal cell structure. The combination of the two can accurately describe its physical properties.

[0109] The gradient intensity distribution of the necrotic region boundary is calculated to assess edge clarity and enhance observation of the boundary characteristics of the necrotic region. Clear edges with high gradient intensity and blurred edges with low gradient intensity correspond to different growth dynamics states. Based on the tumor growth dynamics model, the three-dimensional characteristic parameters of necrotic region size, density, and edge clarity are input into the model. Using a nonlinear mapping algorithm, irregular necrotic regions (large area, low density, blurred edges) are mapped to the rapid growth stage, while regular necrotic regions (small area, high density, clear edges) are mapped to the slow growth stage, thus achieving cross-dimensional inference from morphological characteristics to growth rate.

[0110] This step establishes a cross-scale correlation analysis framework between morphological features and growth dynamics, breaking through the limitations of traditional segmentation, which focuses solely on regional boundaries. By quantifying the three-dimensional morphological parameters of necrotic regions, it is possible to accurately infer tumor growth activity at different stages.

[0111] In the pathological analysis of liver cancer, large areas of low-density necrotic areas with blurred edges often indicate that the tumor is in a rapid proliferation phase with active angiogenesis, while small areas of high-density necrotic areas with clear edges may correspond to the fibrosis degeneration stage after treatment.

[0112] This multi-parameter fusion growth rate estimation method provides clinical evaluation indicators with more biological significance than simple volume measurement, allowing doctors to judge the tumor's response to current treatment and adjust the treatment plan. In practical applications, this mechanism can accurately infer the growth stage differences between primary and metastatic lesions by analyzing the differences in the morphological characteristics of necrotic areas in tumors of different tissue origins. This provides a key imaging basis for tumor heterogeneity research and personalized treatment, demonstrating its unique application value in tumor pathological grading and prognostic assessment.

[0113] like Figure 4 As shown, the relative growth rate of the tumor at different stages is estimated based on the morphological characteristics of the necrotic area, specifically including:

[0114] S310, detecting and segmenting a necrotic area in the multi-dimensional feature cluster;

[0115] S320, calculating the area ratio of the necrotic area and the standard deviation of the internal pixel intensity, and quantifying the size and density of the necrotic area;

[0116] S330, calculating the gradient intensity distribution of the boundary of the necrotic region and evaluating the edge clarity of the necrotic region;

[0117] S340, based on the tumor growth dynamics model, combined with the size, density and edge clarity of the necrotic area, the necrotic area is classified into stages, and an estimate of the relative growth rate of the tumor area is output.

[0118] Large, low-density, irregular necrotic areas with highly blurred edges were mapped as rapid growth stages;

[0119] Necrotic areas with smaller areas, relatively higher density, and clearer edges are mapped as relatively slow or stable growth stages.

[0120] S400, dynamically adjusting the sensitivity threshold of the U-Net decoder segmentation based on the invasion probability information, relative growth rate, and tumor region boundary regularity information output by the invasion direction probability model;

[0121] Increase segmentation sensitivity for areas with rapid growth, blurred boundaries, and irregular necrotic areas;

[0122] Reduce segmentation sensitivity for areas that grow relatively slowly and have regular boundaries;

[0123] The process of dynamically adjusting the U-Net decoder's segmentation sensitivity threshold essentially builds an intelligent control mechanism based on multi-dimensional tumor biological characteristics. This mechanism uses the pseudocapsule breakthrough probability map output by the invasion direction probability model as a spatial risk guide, and identifies high-probability invasion areas (such as pseudocapsule weak points and high-risk areas around blood vessels) as segmentation-sensitive areas that require special attention.

[0124] The relative growth rate obtained from the analysis of the necrotic area is used as the activity indicator in the time dimension, and the tumor area in the rapid growth stage (the lesion corresponding to the large area of ​​low-density necrotic area with blurred edges) is set as the dynamic area where the segmentation sensitivity needs to be enhanced. At the same time, combined with the boundary regularity information of the tumor area (edge ​​clarity evaluated by gradient intensity distribution), the irregular area with blurred boundaries is given a higher segmentation sensitivity adjustment weight.

[0125] The information of these three dimensions is fused through the setting of threshold intervals to form differentiated markers for high-sensitivity demand areas (high probability of invasion + rapid growth + blurred boundaries) and low-sensitivity demand areas (low probability of invasion + slow growth + regular boundaries). Finally, through the adaptive adjustment of the activation function threshold, the decoder can realize targeted segmentation processing of areas with different biological characteristics.

[0126] This step converts biological parameters such as invasion risk, growth rate, and boundary characteristics into sensitivity over-regulation signals of the segmentation model, enabling the decoder to "intelligently" adjust the segmentation strategy according to the actual biological characteristics of the lesion.

[0127] In liver cancer pathology images, dynamically increasing the segmentation sensitivity for tumor edge areas with a high probability of pseudocapsule breakthrough and rapid growth can accurately capture tiny invasive foci and avoid missed diagnosis; while reducing the sensitivity for tumor core areas with slow growth and regular boundaries can effectively suppress noise interference and reduce missegmentation.

[0128] This adaptive regulation mechanism has demonstrated unique value in the pathological analysis of different types of tumors. For example, when processing pancreatic cancer pathological sections with high heterogeneity, the segmentation accuracy can be dynamically adjusted based on the differences in invasive activity in different regions. This allows the segmentation results to not only present the spatial distribution of the tumor, but also reflect its biological behavior characteristics, providing clinical pathological diagnosis with information that combines the dual value of anatomical positioning and biological evaluation.

[0129] This mechanism deeply integrates imaging segmentation with tumor biological characteristics, promoting the leap of pathological image analysis from "morphological description" to "biological understanding", and providing an innovative technical path for precise pathological diagnosis.

[0130] like Figure 5 As shown, the dynamic adjustment of the sensitivity threshold of the U-Net decoder segmentation specifically includes:

[0131] S410, comprehensively analyzing the path probability map, growth rate estimation results, and edge clarity, and setting corresponding recognition threshold intervals respectively.

[0132] S420, based on the threshold interval, read high-probability areas, fast-growing areas, and areas with fuzzy boundaries, and mark them as high-sensitivity demand areas; read low-probability areas, slow-growing areas, and areas with regular boundaries, and mark them as low-sensitivity demand areas;

[0133] S430 : Setting matching activation function thresholds for the high-sensitivity requirement area and the low-sensitivity requirement area respectively.

[0134] S500, using the U-Net decoder with dynamic sensitivity adjustment, segment the input pathological image and output the segmentation result map.

[0135] When using a dynamically sensitive U-Net decoder for pathological image segmentation, bilinear interpolation upsampling is used for areas with high sensitivity requirements (such as those with high invasion probability, rapidly growing areas, and areas with blurred boundaries). This method uses a weighted average of adjacent pixel values ​​to retain more detailed information when magnifying the image. Combined with the nonlinear enhancement of weak signals by the Sigmoid activation function, it can accurately capture tiny invasive lesions or low-contrast tumor boundaries.

[0136] For areas with low sensitivity requirements (low invasion probability, slow growth and regular boundary areas), nearest neighbor upsampling is used to suppress noise interference by retaining the original pixel values, and then the non-specific signals are filtered through the Sigmoid function.

[0137] During the feature fusion stage, the basic visual feature layer and mid-level texture feature layer extracted by the encoder are weighted and concatenated with the decoder's dynamic sensitivity feature map via skip connections. The dynamic feature fusion weight coefficient comprehensively considers the probability of breakthrough of the pseudocapsule, growth rate, and edge clarity of the necrotic area, achieving adaptive feature fusion for regions with different biological characteristics. Regions with high invasion risk are assigned a higher weight for breakthrough probability, thus more closely integrating the encoder's high-level morphological features with the decoder's segmentation features.

[0138] In the final three-channel segmentation map, the tumor core area corresponds to the densely populated cell proliferation area, the pseudocapsular invasion area represents the infiltration boundary of the tumor into the surrounding tissue, and the microvascular invasion area marks the active area of ​​tumor angiogenesis. The three constitute a complete spatial map of tumor biological behavior.

[0139] This step establishes a closed-loop mapping mechanism from "biological characteristics to segmentation strategy to clinical information." Dynamic sensitivity adjustment enables the decoder to perform differentiated segmentation based on the invasive activity and growth status of different tumor regions. For example, in liver tumor pathology images, bilinear interpolation of areas with a high probability of pseudocapsule breakthrough can accurately delineate microinvasion foci, avoiding the boundary loss caused by traditional fixed sampling methods.

[0140] The three-channel segmentation output breaks through the limitations of traditional single-channel segmentation and provides multi-dimensional diagnostic information for clinicians. The precise positioning of the tumor core area can assist in radiotherapy target planning, the extent of the pseudocapsule invasion area can assess the risk of surgical margins, and the distribution of the microvascular invasion area is directly related to the tumor's metastatic potential. This segmentation method, which integrates multi-scale features and biological parameters, has demonstrated unique value in the pathological analysis of different types of tumors (such as glioblastomas with complex vascular networks and invasive breast cancer with variable capsule structures). When dealing with high-grade brain gliomas with large areas of necrosis, dynamic weight adjustment can effectively distinguish between active tumor cells and fibrotic tissue around the necrotic area, so that the segmentation results not only present anatomical boundaries, but also reflect the biological activity gradient of tumor cells.

[0141] This step deeply couples the deep learning segmentation model with the pathophysiological characteristics of tumors, promoting the upgrade of pathological image analysis from "morphological segmentation" to "functional analysis", and providing technical support with both quantitative accuracy and biological significance for accurate tumor diagnosis and treatment decisions.

[0142] like Figure 6 As shown, the U-Net decoder with dynamic sensitivity adjustment is used to segment the input pathological image, specifically including:

[0143] S510 uses bilinear interpolation upsampling for high-sensitivity areas and applies Sigmoid activation function to enhance weak signals;

[0144] S520: Nearest neighbor upsampling is used for low-sensitivity areas, and the Sigmoid activation function is applied to suppress noise.

[0145] S530, weighted concatenation of the basic visual feature layer and the intermediate texture feature layer extracted by the encoder and the dynamic sensitivity feature map of the corresponding layer of the decoder is performed through skip connection, wherein the dynamic feature fusion weight coefficient of the weighted concatenation is W fuse :

[0146] W fuse =α·P break +β·GrowthSpeed+γ·(1-GradSharpness);

[0147] Among them, α, β, γ are weight coefficients, P break is the probability of breaking through the pseudocapsule, GrowthSpeed ​​is the growth speed estimation result, and GradSharpness is a quantitative indicator of the edge clarity of the necrotic area;

[0148] S540, outputting a three-channel segmentation map, including: a tumor core area, a pseudocapsule invasion area, and a microvascular invasion area.

[0149] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0150] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A tumor pathology image segmentation method based on a U-Net neural network, characterized in that: The method comprises: The pathological image is input into the U-Net neural network for encoding, and the grayscale features, texture features, and shape features of the tumor cells in the image are extracted at different levels of the encoder. The hierarchical features are then integrated to construct a multidimensional feature cluster that characterizes the distribution and morphology of cancer cells. Based on the multidimensional feature cluster, a probability model of tumor cell invasion direction is established to quantify the probability of tumor cells breaking through the pseudocapsule and invading the liver parenchyma, and to simulate the potential path of tumor cell spread; In the multidimensional feature cluster, the size, density and edge definition morphological features of the necrotic area in the pathological image are identified and analyzed, and the relative growth rate of the tumor at different stages is estimated based on the morphological features of the necrotic area; Dynamically adjusting the sensitivity threshold of the U-Net decoder segmentation based on the invasion probability information, relative growth rate, and tumor region boundary regularity information output by the invasion direction probability model; The input pathological image is segmented using the U-Net decoder with dynamic sensitivity adjustment, and the segmentation result map is output.

2. The method according to claim 1, characterized in that The construction of a multidimensional feature cluster characterizing the distribution and morphology of cancer cells specifically includes: Using the convolutional and pooling layers of the U-Net encoder, the basic visual features of the image, including pixel grayscale value distribution and local contrast information, are extracted in the first 1-2 downsampling levels of the encoder to form the basic visual feature layer; At the 3rd and 4th downsampling levels of the encoder, the contrast of the gray-level co-occurrence matrix and the local binary pattern histogram are calculated to extract the intermediate texture feature layer describing the density and structural heterogeneity of tumor cells. At encoder level 5 and above, high-level morphological feature layers of tumor cell clusters are extracted through semantic segmentation mask generation and contour fitting; Through jump connections, the basic visual feature layer, the intermediate texture feature layer and the high-level morphological feature layer are channel-joined and fused with 1×1 convolution to generate a multidimensional feature cluster containing multi-scale spatial information and semantic information.

3. The method according to claim 2, characterized in that The method of quantifying the probability of tumor cells breaking through the pseudocapsule and invading the liver parenchyma and simulating the potential path of tumor cell spread specifically includes: Define the pseudocapsule area feature template and the perivascular area feature template; Calculating the matching degree between the multidimensional feature cluster and the pseudocapsule region feature template to generate a probability map of tumor cells breaking through the pseudocapsule; Calculating the matching degree between the multi-dimensional feature cluster and the feature template of the perivascular region, and combining the vascular direction information to generate a path probability map of tumor cells spreading along the venous wall; The pseudocapsule breakthrough probability map and the path probability map are integrated to form a comprehensive tumor invasion direction probability model output.

4. The method according to claim 3, characterized in that The matching degree is calculated to generate a probability map of tumor cells breaking through the pseudocapsule, specifically: The sliding window is traversed to calculate the matching degree between the multidimensional feature cluster at each position and the feature template of the pseudo-encapsulation area: Among them, MatchScore(x,y) represents the matching degree between the multidimensional feature cluster at position (z,y) and the feature template of the pseudocapsule area, F cluster (x,y) represents the feature cluster vector at position (x,y), T capsule Represents the characteristic template of the pseudocapsule area; Normalize the matching degree: Among them, min(Score) represents the minimum value of the matching degree between the multidimensional feature cluster and the pseudocapsule region feature template, and max(Score) represents the maximum value of the matching degree between the multidimensional feature cluster and the pseudocapsule region feature template; Output P break (x,y) is used as the probability map of breaking through the false envelope.

5. The method according to claim 3, characterized in that The matching degree calculation is performed, and combined with the blood vessel direction information, to generate a path probability map of tumor cells spreading along the vein wall, specifically including: Among them, VascMatch(x,y) is the matching degree between the multidimensional feature cluster at position (x,y) and the feature template of the perivascular area, T vessel is the feature template of the perivascular area, σ is the Gaussian kernel width; Extract the vascular skeleton line direction field θ(x,y) and generate the direction weight map W direction (x,y): W direction (x,y)=cos 2 (φ(x,y)-θ(x,y)); Where φ(x,y) represents the link direction from position (x,y) to the centerline of the tumor; Fusion generates path probability graph P spread (x,y): P spread (x,y)=VascMatch(x,y)×W direction (x,y)。 6. The method according to claim 3, characterized in that The inferring the relative growth rate of the tumor at different stages based on the morphological characteristics of the necrotic area specifically includes: Detecting and segmenting necrotic regions in the multidimensional feature cluster; Calculating the area ratio of the necrotic area and the standard deviation of the internal pixel intensity to quantify the size and density of the necrotic area; Calculating the gradient intensity distribution of the boundary of the necrotic area to evaluate the edge clarity of the necrotic area; Based on the tumor growth dynamics model, the necrotic area is classified into stages in combination with the size, density and edge clarity of the necrotic area, and the relative growth rate estimation of the tumor area is output.

7. The method according to claim 6, characterized in that The dynamic adjustment of the sensitivity threshold of the U-Net decoder segmentation specifically includes: Comprehensively analyze the path probability map, growth rate estimation results, and edge clarity, and set corresponding recognition threshold intervals respectively. Based on the threshold interval, high-probability areas, fast-growing areas, and areas with fuzzy boundaries are read and marked as high-sensitivity demand areas; low-probability areas, slow-growing areas, and areas with regular boundaries are read and marked as low-sensitivity demand areas; Matching activation function thresholds are set for the high-sensitivity requirement area and the low-sensitivity requirement area respectively.

8. The method according to claim 7, characterized in that The method of segmenting the input pathological image using the U-Net decoder with dynamic sensitivity adjustment specifically includes: Bilinear interpolation upsampling is used for high-sensitivity areas, and Sigmoid activation function is applied to enhance weak signals; For low-sensitivity areas, the nearest neighbor upsampling is used, and the Sigmoid activation function is applied to suppress noise; The basic visual feature layer and intermediate texture feature layer extracted by the encoder are weightedly spliced ​​with the dynamic sensitivity feature map of the corresponding layer of the decoder through jump connections, where the dynamic feature fusion weight coefficient of the weighted splicing is W fuse : W fuse =α·P break +β·GrowthSpeed+γ·(1-GradSharpness); Among them, α, β, γ are weight coefficients, P break is the probability of breaking through the pseudocapsule, GrowthSpeed ​​is the growth speed estimation result, and GradSharpness is a quantitative indicator of the edge clarity of the necrotic area; Output a three-channel segmentation map, which includes: tumor core area, pseudocapsule invasion area and microvascular invasion area.

Images