Tumor budding automatic identification method fusing pathological staining migration and YOLO target detection

Abstract

The invention provides an automatic tumor budding identification method fusing pathological staining migration and YOLO target detection. The method comprises the following steps: training Hamp by using a CUT model; a staining migration model is established from the E image to the immunohistochemical image, and the staining migration model is utilized to carry out Hamp; e, converting the dyed full-width pathological section into a cytokeratin dyed full-width pathological section; and performing tumor budding detection on the generated full-width pathological section dyed by the cytokeratin by using a TB-YOLO model improved on the basis of YOLOv10. According to the method, the CUT is combined for pathological image staining migration, the IHC image generated through TB-YOLO analysis is used for tumor budding detection, meanwhile, accurate translation and detection of positive signals are ensured, richer diagnosis information is provided for pathologists, the diagnosis accuracy is improved, the pathological examination cost of patients is reduced, and the purpose that only Hamp is used for diagnosis is achieved. And E, the tumor budding is accurately detected by the dyed image, so that the comparability of medical images in different dyeing modes and the applicability of a cross-dyeing method are improved.

Description

Technical Field

[0001] The present invention relates to the field of intersectional technology of medical artificial intelligence and digital pathology, and more specifically, to an automatic tumor budding recognition method integrating pathological staining migration and YOLO target detection. Background Art

[0002] In medical pathology diagnosis, H&E (hematoxylin-eosin) staining is a widely used routine staining method for pathological tissues, which can clearly present the cell arrangement and tissue morphological characteristics. However, H&E staining cannot specifically identify molecular markers in the tumor microenvironment (such as cytokeratin CK), which leads to significant limitations in tumor typing, invasiveness assessment and prognosis judgment. In contrast, IHC (immunohistochemistry) staining can accurately locate tumor buds (TBs) by targeting specific protein markers (such as CK), thereby effectively circumventing the shortcomings of H&E staining. However, the production cost of IHC staining is high, the experimental cycle is long, and it relies on professional laboratory conditions, making it difficult to popularize in small and medium-sized medical institutions. In addition, tumor budding detection itself also faces many challenges. For example, the budding area is usually small in size and irregular in shape, with blurred boundaries with surrounding normal tissues. In addition, the noise and artifact interference during the staining process make traditional detection methods prone to missed detection or misjudgment, seriously affecting the accuracy of diagnosis.

[0003] Conventional approaches to automated tumor budding detection in IHC pathology images typically involve the following steps: First, the operator manually delineates a region of interest (ROI) at a specific anatomical site, such as the tumor invasion front, based on pathological experience, while manually excluding interfering areas containing necrotic tissue. Second, within the selected ROI, threshold-based image segmentation combined with morphological filtering, or supplemented with a shallow convolutional neural network, is used to extract cell nuclei features. Finally, tumor budding is screened using a preset size threshold. For example, Fauzi et al. developed an automated tumor budding detection method using cytokeratin AE1 / 3-stained colorectal cancer WSIs. The method uses thresholding and morphological operations to detect tumor boundaries and employs an Otsu threshold to distinguish tumor nuclei from background for tissue segmentation. Fischer et al. used Qupath software to semi-automatically assess tumor budding in colorectal cancers stained with CK. However, these methods often detect more tumor buds than manual counting, potentially leading to overestimation of the number. Weis et al. used a CNN model to automatically detect tumor buds in colorectal cancer slides and performed post-processing operations to improve the accuracy of TB detection. In addition, the CUT (Contrastive Unpaired Translation) model is an unsupervised image translation model that can produce more natural and realistic image translation results without paired images. The YOLOv10 model is a deep learning-based object detection model that inherits the fast and efficient features of the YOLO series and optimizes its architecture, feature extraction, and detection performance.

[0004] Although existing technologies can achieve the detection of tumor budding to a certain extent, there are still many defects. First, traditional detection methods rely on manual delineation of ROI and threshold segmentation, which is not only time-consuming and labor-intensive, but also requires extremely high experience from the operator, and is prone to inaccurate detection results due to subjective factors. Secondly, existing technologies have poor adaptability to noise and morphological changes, and it is difficult to effectively deal with noise and artifact interference during the staining process, resulting in insufficient robustness of detection. In addition, when processing complex pathological images, existing technologies are often unable to effectively capture the subtle differences between tumor budding cells and surrounding tissues, resulting in low detection accuracy. At the same time, the high cost and long cycle of IHC staining limit its widespread application in clinical practice, and although H&E staining is low-cost and easy to operate, it lacks the ability to specifically detect tumor budding. Therefore, the development of a technology that can accurately convert H&E-stained pathological images into CK-stained IHC pathological images and perform tumor budding detection on virtual CK-stained pathological images has significant medical application value for improving the accuracy of pathological diagnosis, optimizing clinical decision-making, and reducing detection costs. Summary of the Invention

[0005] To address the aforementioned technical issues, we provide a method for automatically identifying tumor budding that integrates pathological staining migration and YOLO object detection. This method uses a CUT model to accurately convert H&E-stained images to IHC-stained images, and an improved TB-YOLO model to accurately detect tumor budding.

[0006] The technical means adopted in the present invention are as follows:

[0007] An automatic tumor budding recognition method integrating pathological staining migration and YOLO target detection includes:

[0008] S1. Use the CUT model to train a staining migration model between H&E images and immunohistochemistry (IHC) images, and use the staining migration model to convert H&E-stained whole-frame pathology sections (WSIs) into cytokeratin (CK)-stained whole-frame pathology sections (WSIs);

[0009] S2. Tumor budding detection was performed on the generated cytokeratin (CK) stained whole-frame pathology sections (WSIs) using the improved TB-YOLO model based on YOLOv10.

[0010] Furthermore, step S1 specifically includes:

[0011] S11. The generator (G) converts the source domain X image into a stylized target domain Y image. The discriminator (D) distinguishes the generated image from the real target domain image through adversarial loss, driving the generator to improve the output realism.

[0012] S12. Introducing a self-supervised contrastive learning mechanism, the generator extracts multi-level local features of the source image as anchors, and uses the features at the corresponding positions of the generated image as positive samples. At the same time, features at other positions of the same image or other images in the same batch are used as negative samples. Through a multi-layer feature contrast loss, the generated image is forced to implicitly align with the source image in the semantically key areas.

[0013] In this embodiment, multi-scale feature map extraction is used to achieve collaborative optimization of global structure and local details, while adversarial loss ensures that the style of the generated image matches the target domain distribution. The contrast loss and adversarial loss are jointly optimized to achieve high-fidelity cross-domain image conversion.

[0014] Furthermore, in step S2, the TB-YOLO model integrates the SEAttention module, the C3k2 module, the C2f_EMA module, and the Detect_FASF module to improve feature extraction and target detection performance, wherein:

[0015] The SEAttention module introduces a channel attention mechanism to enhance the expressiveness of key features;

[0016] The C3k2 module uses a dynamic kernel mechanism to capture local and global context information, improving detection accuracy while preserving spatial details;

[0017] The C2f_EMA module optimizes the structure of the C2f module, enhances feature representation capabilities and retains multi-scale information;

[0018] The Detect_FASF module reduces cross-scale feature loss and ensures robust fusion of features between different resolutions.

[0019] Furthermore, in step S2, the TB-YOLO model includes a backbone network, a feature fusion network, and a detection head, wherein:

[0020] The backbone network includes multiple convolution operations and feature extraction modules;

[0021] The feature fusion network fuses feature maps of different scales through upsampling and splicing operations;

[0022] The detection head fuses feature maps of different scales through the Detect_FASF module to generate multi-scale detection results.

[0023] Furthermore, in the backbone network:

[0024] The input image first undergoes a convolution operation to generate the first layer feature map P1;

[0025] The first-layer feature map P1 undergoes one convolution to generate the second-layer feature map P2, which then enters the feature extraction unit consisting of three C2f modules;

[0026] The second-layer feature map P2 undergoes convolution operation to generate the third-layer feature map P3, and the features are further optimized through 6 C2f modules;

[0027] The third-layer feature map P3 passes through the SCDown module to generate the fourth-layer feature map P4, which then enters the C2f module six times;

[0028] The fourth-layer feature map P4 is converted into the fifth-layer feature map P5 through the SCDown module, and the feature representation is optimized through the C2fCIB module three times;

[0029] The fifth-layer feature map P5 finally passes through the SPPF module and SEAttention module to obtain high-level semantic features.

[0030] Furthermore, in the feature fusion network:

[0031] The fifth layer feature map P5 is scaled by 2 times through upsampling and concatenated with the fourth layer feature map P4;

[0032] The concatenated features are input into the C3k2 module to further extract deep features;

[0033] The processed feature map is upsampled again, concatenated with the third-layer feature map P3, and high-precision features are generated through the C3k2 module;

[0034] Upsample the concatenated features to the size of the second-layer feature map P2 and fuse them with the second-layer feature map P2;

[0035] The fused features are input into the C2f_EMA module to further optimize the features;

[0036] The features of the second-layer feature map P2, the third-layer feature map P3, the fourth-layer feature map P4, and the fifth-layer feature map P5 are subjected to multiple up- and down-sampling and splicing operations, so that feature maps of all scales are optimized by the C2f_EMA module.

[0037] Furthermore, in step S2, the process of tumor budding detection includes:

[0038] S21, downsampling IHC staining WSI to generate thumbnails for panoramic visualization analysis;

[0039] S22. Perform color deconvolution on the thumbnail to separate the hematoxylin (H), eosin (E), and 3,3'-diaminobenzidine (DAB) staining channels, where the DAB channel specifically marks the CK staining-positive expression area;

[0040] S23, the DAB channel is processed by combining Otsu threshold segmentation and morphological closing operation to achieve preliminary binary detection of the tumor area;

[0041] S24. Connected domain analysis was used to filter out small fragmented regions and retain the continuous tumor region.

[0042] S25. Based on the biological characteristics of the tumor-stroma interface, the detection area was expanded by 500 μm to cover the tumor invasion front.

[0043] S26, using a sliding window strategy to densely sample the invasion front area and generate high-resolution image patches to input into the pre-trained TB-YOLO model;

[0044] S27, converting the coordinates of the tumor bud detected in each image block to the original WSI coordinate system through coordinate mapping, and storing the structured detection results in JSON format, including the target location, confidence, and bounding box parameters;

[0045] S28. The detection results are superimposed on the original WSI through the digital pathology visualization platform, where high-confidence tumor buds with a confidence level > 0.9 are explicitly marked in the form of geometric annotations to generate a quantifiable pathology report that complies with the DICOM standard.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] 1. The present invention provides an automatic tumor budding identification method that integrates pathological staining migration and YOLO target detection. It can learn the mapping relationship between the H&E image domain and the IHC image domain in an unsupervised manner, and accurately convert the H&E staining WSI into the CK staining WSI. This greatly improves the comparability of medical images under different staining methods and the applicability across staining methods, reduces the dependence on IHC staining, and reduces the cost of pathological examination.

[0048] 2. This paper provides an automated tumor budding detection method that integrates pathological stain migration and YOLO object detection, significantly improving the accuracy, robustness, and efficiency of tumor budding detection. Through multi-scale feature fusion and optimized detection head design, the TB-YOLO model can effectively detect tumor budding cells of varying sizes, improving detection accuracy and completeness, and providing a more reliable basis for pathological diagnosis.

[0049] 3. The present invention provides an automatic tumor budding identification method that integrates pathological staining migration and YOLO target detection, which achieves efficient display and structured storage of test results, facilitating pathologists to quickly and accurately view and analyze test results, thereby improving diagnostic efficiency and accuracy. Furthermore, the DICOM-compliant report format facilitates sharing and communication between different medical systems.

[0050] 4. The present invention provides an automatic tumor budding identification method that integrates pathological staining migration and YOLO target detection, which enables the model to demonstrate good generalization ability on different pathological image datasets, reduces the dependence on large-scale paired annotated data, reduces the cost and time of data preparation, and improves the practicality and scalability of the model.

[0051] The present invention not only enables accurate detection of tumor budding using only H&E-stained pathological sections, but also greatly improves the comparability of medical images under different staining methods and the applicability of cross-staining methods, reduces the cost of pathological examination, improves diagnostic efficiency and accuracy, and provides strong support for pathological diagnosis and clinical decision-making.

[0052] Based on the above reasons, the present invention can be widely promoted in the fields of medical artificial intelligence and digital pathology. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0054] Figure 1 Flow chart of the method of the present invention.

[0055] Figure 2 This is the architecture diagram of the CUT model of the present invention.

[0056] Figure 3 This is a training flow chart of the CUT model of the present invention.

[0057] Figure 4 This is the architecture diagram of the TB-YOLO network model of the present invention.

[0058] Figure 5 This is a training flow chart of the TB-YOLO network model of the present invention.

[0059] Figure 6 This is a flowchart of the H&E-to-IHC domain migration provided by an embodiment of the present invention.

[0060] Figure 7 A flowchart of tumor budding detection based on generated IHC images provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0062] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or apparatuses.

[0063] like Figure 1 As shown, the present invention provides a method for automatically identifying tumor budding by integrating pathological staining migration and YOLO target detection, comprising:

[0064] S1. Use the CUT model to train a staining migration model between H&E images and immunohistochemistry (IHC) images, and use the staining migration model to convert H&E-stained whole-frame pathology sections (WSIs) into cytokeratin (CK)-stained whole-frame pathology sections (WSIs);

[0065] S2. Tumor budding detection was performed on the generated cytokeratin (CK) stained whole-frame pathology sections (WSIs) using the improved TB-YOLO model based on YOLOv10.

[0066] In specific implementation, as a preferred embodiment of the present invention, step S1 specifically includes:

[0067] S11. The generator (G) converts the source domain X image into a stylized target domain Y image. The discriminator (D) distinguishes the generated image from the real target domain image through adversarial loss, driving the generator to improve the output realism.

[0068] S12. Introducing a self-supervised contrastive learning mechanism, the generator extracts multi-level local features of the source image as anchors, and uses the features at the corresponding positions of the generated image as positive samples. At the same time, features at other positions of the same image or other images in the same batch are used as negative samples. Through a multi-layer feature contrast loss, the generated image is forced to implicitly align with the source image in the semantically key areas.

[0069] In this embodiment, the CUT model is an image style transfer framework based on unsupervised adversarial learning, whose core consists of a generator (G) and a discriminator (D). The innovation of the CUT model lies in the introduction of a self-supervised contrastive learning mechanism: during the training process, the generator first extracts the multi-level local features of the source image as anchors, and uses the features of the corresponding positions of the generated image as positive samples (such as Figure 2 Medium blue box + As shown), the features of other locations in the same image or other images in the same batch are used as negative samples (as shown Figure 2 Yellow box z -As shown). Through multi-layer feature contrast loss (such as InfoNCE loss), the model forces the generated image to be implicitly aligned with the source image in semantically key areas (such as object edges, texture structures) without relying on paired supervision data. Specifically, the goal of the contrast loss is to maximize the similarity between the anchor point and the positive sample, while minimizing the similarity between the anchor point and the negative sample, thereby preserving content consistency. In addition, CUT achieves collaborative optimization of global structure and local details through multi-scale feature map extraction, such as sampling local blocks from feature maps at different levels of the generator, while the adversarial loss ensures that the style of the generated image matches the target domain distribution. By jointly optimizing the contrast loss and the adversarial loss, CUT achieves high-fidelity cross-domain image conversion under unsupervised conditions, which can not only maintain the content structure of the source image, but also accurately migrate the target domain style features. As shown Figure 3 FIG. 1 is a training flow chart of the CUT model of the present invention.

[0070] In specific implementation, as a preferred embodiment of the present invention, in step S2, the TB-YOLO model integrates the SEAttention module, the C3k2 module, the C2f_EMA module and the Detect_FASF module to improve feature extraction and target detection performance, wherein:

[0071] The SEAttention module introduces a channel attention mechanism to enhance the expressiveness of key features;

[0072] The C3k2 module uses a dynamic kernel mechanism to capture local and global context information, improving detection accuracy while preserving spatial details;

[0073] The C2f_EMA module optimizes the structure of the C2f module, enhances feature representation capabilities and retains multi-scale information;

[0074] The Detect_FASF module reduces cross-scale feature loss and ensures robust fusion of features between different resolutions.

[0075] In this embodiment, this design enables the TB-YOLO model to effectively detect objects of different sizes, improving detection accuracy and robustness. Through the synergistic effect of the above modules, the TB-YOLO model achieves efficient feature extraction and multi-scale object detection capabilities, enabling detection in complex scenes.

[0076] In specific implementation, as a preferred embodiment of the present invention, in step S2, the TB-YOLO model includes a backbone network, a feature fusion network, and a detection head, wherein:

[0077] The backbone network includes multiple convolution operations and feature extraction modules;

[0078] The feature fusion network fuses feature maps of different scales through upsampling and splicing operations;

[0079] The detection head fuses feature maps of different scales through the Detect_FASF module to generate multi-scale detection results.

[0080] When specifically implemented, as a preferred embodiment of the present invention, Figure 4 As shown, in the backbone network:

[0081] The input image first undergoes a convolution operation (kernel size 3, stride 2, number of channels 64) to generate the first layer feature map P1;

[0082] The first-layer feature map P1 undergoes a convolution (kernel size 3, stride 2, number of channels 128) to generate the second-layer feature map P2, which then enters the feature extraction unit consisting of three C2f modules;

[0083] The second-layer feature map P2 undergoes a convolution operation (kernel size 3, stride 2, number of channels 256) to generate the third-layer feature map P3, and the features are further optimized through 6 C2f modules;

[0084] The third-layer feature map P3 passes through the SCDown module (kernel size 3, step size 2, number of channels 512) to generate the fourth-layer feature map P4, which then enters the C2f module 6 times;

[0085] The fourth-layer feature map P4 is processed by the SCDown module (kernel size 3, stride 2, number of channels 1024) to generate the fifth-layer feature map P5, and the feature representation is optimized by the C2fCIB module three times;

[0086] The fifth-layer feature map P5 finally passes through the SPPF module (kernel size is 5) and the SEAttention module (channel scaling rate is 16) to obtain high-level semantic features.

[0087] In specific implementation, as a preferred embodiment of the present invention, in the feature fusion network:

[0088] The fifth layer feature map P5 is scaled by 2 times through upsampling (using nearest neighbor interpolation) and feature concatenated with the fourth layer feature map P4;

[0089] The concatenated features are input into the C3k2 module to further extract deep features;

[0090] The processed feature map is upsampled again, concatenated with the third-layer feature map P3, and high-precision features are generated through the C3k2 module;

[0091] Upsample the concatenated features to the size of the second-layer feature map P2 and fuse them with the second-layer feature map P2;

[0092] The fused features are input into the C2f_EMA module to further optimize the features;

[0093] The features of the second layer feature map P2, the third layer feature map P3, the fourth layer feature map P4 and the fifth layer feature map P5 are subjected to multiple up- and down-sampling and splicing operations, so that the feature maps of all scales are optimized by the C2f_EMA module. Figure 5 FIG. 1 is a training flow chart of the TB-YOLO network model of the present invention.

[0094] When specifically implemented, as a preferred embodiment of the present invention, Figure 6 This paper demonstrates a cross-modal transfer process from H&E-stained WSI to the IHC domain. First, H&E-stained WSI and IHC-stained WSI are spatially registered using the DeepHistReg algorithm to generate geometrically aligned IHC-stained WSI. Subsequently, local image patches are extracted from the registered H&E-IHC image pairs to construct a training dataset. Network training utilizes a generative adversarial framework, extracting multi-level features through the generator encoder. This is combined with adversarial supervision from the discriminator and a multi-layer patchwise contrastive loss to achieve cross-modal feature alignment. This mechanism achieves pathology-preserving domain transfer by layer-by-layer decoupling the domain differences between H&E and IHC images in tissue topology (e.g., glandular distribution) and subcellular texture (e.g., chromatin granularity). The trained H&E-to-IHC model can infer IHC-style image patches from the input H&E-stained WSI. Virtual IHC-stained WSIs are then synthesized using a panoramic reconstruction algorithm (WSI stitching). This method achieves high fidelity cross-modal conversion of pathological images while maintaining key diagnostic features such as cell nuclear morphology and chromatin pattern through the collaborative optimization of adversarial learning and contrastive learning, providing a reliable data basis for quantitative pathology analysis.

[0095] In specific implementation, as a preferred embodiment of the present invention, in step S2, as Figure 7 As shown, the process of tumor budding detection includes:

[0096] S21, downsampling IHC staining WSI to generate thumbnails for panoramic visualization analysis;

[0097] S22. Perform color deconvolution on the thumbnail to separate the hematoxylin (H), eosin (E), and 3,3'-diaminobenzidine (DAB) staining channels, where the DAB channel specifically marks the CK staining-positive expression area;

[0098] S23, the DAB channel is processed by combining Otsu threshold segmentation and morphological closing operation to achieve preliminary binary detection of the tumor area;

[0099] S24. Connected domain analysis was used to filter out small fragmented regions and retain the continuous tumor regions with pathological significance.

[0100] S25: Based on the biological characteristics of the tumor-stroma interface, the detection area was expanded by 500 μm to cover the tumor invasion front.

[0101] S26: A sliding window strategy (window size 1280×1280 pixels, 40× magnification, step size 640 pixels, 50% overlap) was used to densely sample the invasion front area and generate high-resolution image patches to input into the pre-trained TB-YOLO model.

[0102] S27, converting the coordinates of the tumor bud detected in each image patch to the original WSI coordinate system through coordinate mapping, and storing the structured detection results in JSON format, including the target location, confidence, and bounding box parameters;

[0103] S28. The detection results are superimposed on the original WSI through a digital pathology visualization platform, where high-confidence tumor buds with a confidence level > 0.9 are explicitly marked in the form of geometric annotations (such as polygonal / rectangular ROIs) to generate a quantifiable pathology report that complies with the DICOM standard.

[0104] Example 1

[0105] The patient went to a primary hospital for treatment due to blood in the stool and changes in bowel habits, and was initially diagnosed with colorectal cancer. After the operation, the pathology department performed routine H&E staining on the resected tissue, but due to equipment limitations, the primary hospital was unable to carry out CK staining to assess the tumor budding status (an internationally recognized prognostic indicator). Traditional CK staining requires special reagents, high-precision equipment and the experience of pathologists, while H&E staining can easily lead to blurred cell boundaries due to differences in the standardization of the staining process, and there are significant subjective deviations in manual judgment of budding. By applying the automatic tumor budding recognition method designed by the present invention, the budding status of colorectal cancer patients can be automatically and accurately judged based on H&E images, accurately matching the adjuvant chemotherapy recommendations based on the budding status in the NCCN guidelines, and providing primary hospitals with low-cost, standardized, and accurate diagnosis solutions.

[0106] Example 2

[0107] After patients with space-occupying lesions found by colonoscopy receive an initial H&E pathological diagnosis, further CK staining is required to assess the tumor invasion pattern in order to develop a personalized plan. However, the cost of CK staining is over 800 yuan per time (including CDX2 / CK20 antibody combinations), and the preparation time is as long as 72 hours. In addition, the inter-observer variability of manual counting of buds is large, which can easily lead to over- or under-treatment. By utilizing the fusion staining migration and tumor budding automatic identification method of the present invention, corresponding IHC slices can be generated based on existing H&E slices in a short time, and budding status reports can be output with high diagnostic consistency, which reduces the comprehensive testing cost for patients.

[0108] In summary, the present invention can convert a patient's existing H&E-stained pathology images into CK IHC-stained pathology images, thereby assisting patients in obtaining IHC-stained sections in a cost-effective and efficient manner. Furthermore, the generated IHC images can be used for tumor budding detection, enabling pathologists to more comprehensively observe the immune microenvironment of colorectal cancer and provide a more reliable basis for precise diagnosis and treatment. The implementation of the present invention includes the following steps: first, constructing a local dataset based on hospital-based H&E-stained images and corresponding CK IHC-stained images; second, using this dataset to train a model designed specifically for pathology image staining transfer; and finally, using the trained model to process the patient's H&E-stained digital pathology slides to generate corresponding IHC-stained pathology images. Pathologists can use the generated virtual IHC-stained pathology images, combined with TB-YOLO, to further analyze the degree of budding in the patient's colorectal cancer. Based on the analysis results, doctors can formulate optimal treatment and prevention plans, such as selecting appropriate chemotherapy drugs and determining whether the patient is suitable for targeted therapy or immunotherapy, thereby optimizing personalized diagnosis and treatment strategies.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatically identifying tumor budding by integrating pathological staining migration and YOLO target detection, characterized in that: include: S1. Use the CUT model to train a staining migration model between H&E images and immunohistochemistry images, and use the staining migration model to convert H&E-stained full-length pathology sections into cytokeratin-stained full-length pathology sections. S2. Use the improved TB-YOLO model based on YOLOv10 to perform tumor budding detection on the generated cytokeratin-stained full-length pathological sections.

2. The method for automatically identifying tumor budding by integrating pathological staining migration and YOLO target detection according to claim 1, characterized in that: Step S1 specifically includes: S11, the generator converts the source domain X image into the target domain Y stylized image, and the discriminator distinguishes the generated image from the real target domain image through adversarial loss, driving the generator to improve the output realism; S12. A self-supervised contrastive learning mechanism is introduced. The generator extracts multi-level local features of the source image as anchor points, and uses the features at the corresponding positions of the generated image as positive samples. At the same time, the features at other positions of the same image or other images in the same batch are used as negative samples. Through multi-layer feature contrast loss, the generated image is forced to be implicitly aligned with the source image in the semantic key area.

3. The method for automatically identifying tumor budding by integrating pathological staining migration and YOLO target detection according to claim 1, characterized in that: In step S2, the TB-YOLO model integrates the SEAttention module, the C3k2 module, the C2f_EMA module, and the Detect_FASF module to improve feature extraction and object detection performance, where: The SEAttention module introduces a channel attention mechanism to enhance the expressiveness of key features; The C3k2 module uses a dynamic kernel mechanism to capture local and global context information, improving detection accuracy while preserving spatial details; The C2f_EMA module optimizes the structure of the C2f module, enhances feature representation capabilities and retains multi-scale information; The Detect_FASF module reduces cross-scale feature loss and ensures robust fusion of features between different resolutions.

4. The method for automatically identifying tumor budding by integrating pathological staining migration and YOLO target detection according to claim 1, characterized in that: In step S2, the TB-YOLO model includes a backbone network, a feature fusion network, and a detection head, wherein: The backbone network includes multiple convolution operations and feature extraction modules; The feature fusion network fuses feature maps of different scales through upsampling and splicing operations; The detection head fuses feature maps of different scales through the Detect_FASF module to generate multi-scale detection results.

5. The method for automatically identifying tumor budding by integrating pathological staining migration and YOLO target detection according to claim 4, characterized in that: In the backbone network: The input image first undergoes a convolution operation to generate the first layer feature map P1; The first-layer feature map P1 undergoes one convolution to generate the second-layer feature map P2, which then enters the feature extraction unit consisting of three C2f modules; The second-layer feature map P2 undergoes convolution operation to generate the third-layer feature map P3, and the features are further optimized through 6 C2f modules; The third-layer feature map P3 passes through the SCDown module to generate the fourth-layer feature map P4, which then enters the C2f module six times; The fourth-layer feature map P4 is converted into the fifth-layer feature map P5 through the SCDown module, and the feature representation is optimized through the C2fCIB module three times; The fifth-layer feature map P5 finally passes through the SPPF module and SEAttention module to obtain high-level semantic features.

6. The method for automatically identifying tumor budding by integrating pathological staining migration and YOLO target detection according to claim 4, characterized in that: In the feature fusion network: The fifth layer feature map P5 is scaled by 2 times through upsampling and concatenated with the fourth layer feature map P4; The concatenated features are input into the C3k2 module to further extract deep features; The processed feature map is upsampled again, concatenated with the third-layer feature map P3, and high-precision features are generated through the C3k2 module; Upsample the concatenated features to the size of the second-layer feature map P2 and fuse them with the second-layer feature map P2; The fused features are input into the C2f_EMA module to further optimize the features; The features of the second-layer feature map P2, the third-layer feature map P3, the fourth-layer feature map P4, and the fifth-layer feature map P5 are subjected to multiple up- and down-sampling and splicing operations, so that feature maps of all scales are optimized by the C2f_EMA module.

7. The method for automatically identifying tumor budding by integrating pathological staining migration and YOLO target detection according to claim 1, characterized in that: In step S2, the process of tumor budding detection includes: S21, downsampling IHC staining WSI to generate thumbnails for panoramic visualization analysis; S22, performing color deconvolution on the thumbnail to separate the hematoxylin, eosin, and 3,3'-diaminobenzidine staining channels, wherein the 3,3'-diaminobenzidine staining channel specifically marks the CK staining-positive expression area; S23, Otsu threshold segmentation and morphological closing operation are combined to process the 3,3'-diaminobenzidine staining channel to achieve preliminary binary detection of the tumor area; S24. Connected domain analysis was used to filter out small fragmented regions and retain the continuous tumor region. S25. Based on the biological characteristics of the tumor-stroma interface, the detection area was expanded by 500 μm to cover the tumor invasion front. S26, using a sliding window strategy to densely sample the invasion front area and generate high-resolution image patches to input into the pre-trained TB-YOLO model; S27, converting the coordinates of the tumor bud detected in each image block to the original WSI coordinate system through coordinate mapping, and storing the structured detection results in JSON format, including the target location, confidence, and bounding box parameters; S28. The detection results are superimposed on the original WSI through the digital pathology visualization platform, where high-confidence tumor buds with a confidence level > 0.9 are explicitly marked in the form of geometric annotations to generate a quantifiable pathology report that complies with the DICOM standard.

Images