A breast cancer pathological image multi-class tissue segmentation method

Abstract

This invention belongs to the field of tissue recognition technology and provides a multi-class tissue segmentation method for breast cancer pathological images, including: pathological tissue image acquisition and multi-class tissue segmentation model recognition; the model pre-training process includes: constructing an image segmentation model that integrates the improved DenseASPP module and the boundary-aware dynamic weighted fusion mechanism, dataset construction, and model training; this invention enhances the multi-scale and global contextual information modeling capability by improving the DenseASPP module, alleviating feature loss caused by downsampling and dilated convolution, and improving the completeness and continuity of region prediction; the dynamic weighted fusion mechanism enhances the discriminative power at the boundaries of various tissues, improves boundary detail perception and regional semantic consistency, and reduces boundary blurring and adhesion phenomena.

Description

Technical Field

[0001] This invention relates to the field of tissue identification technology, and in particular to a method for multi-category tissue segmentation in breast cancer pathological images. Background Technology

[0002] Tumor-infiltrating lymphocytes (TILs) are abbreviated as TILs, artificial intelligence (AI) as AI, and deep learning (DL) as DL. The breast cancer tumor microenvironment is a complex dynamic system composed of tumor epithelial cells, TILs, other immune cells, inflammatory cells, stromal components, and blood vessels and lymphatic vessels. These cells and tissue components form a complex regulatory network through various signaling mechanisms and interactions, playing a crucial role in the occurrence, development, invasion, metastasis, treatment response, and clinical prognosis of breast cancer. Therefore, accurate localization, identification, and segmentation of multiple tissue components in breast cancer pathological images are urgently needed for quantitative analysis of the breast cancer tumor microenvironment in current research and clinical applications. Breast cancer segmentation is a key step in breast medical image analysis, its main task being to accurately extract tumor or suspicious lesion regions from ultrasound, mammography, MRI, or pathological images. To achieve accurate identification of lesion regions, researchers have proposed traditional methods such as threshold segmentation, region segmentation, contour extraction, edge detection, and region growing. In breast cancer image analysis, DL can perform end-to-end modeling of image data, identifying and evaluating potentially malignant lesion regions, thereby supporting the diagnosis and segmentation of breast cancer. In the field of medical image segmentation, DeepLabv3, as one of the classic deep learning models, excels in expanding the receptive field and enhancing multi-scale feature representation by leveraging dilated convolution and multi-scale contextual modeling mechanisms, providing an important foundation for semantic segmentation in complex pathological image scenarios. It has become a leader in automatic breast lesion segmentation. In recent years, research in medical image segmentation has focused on the automatic segmentation of cell nuclei, cellular tissue regions, and cancerous structures. Overall, most existing methods are based on the classic semantic segmentation framework, with targeted improvements made to address the complex tissue structures, significant staining differences, and blurred boundaries in pathological images. Researchers have attempted to introduce mechanisms such as dual attention, self-attention nested paths, EPO optimization strategies, Unet-like architectures, and multi-task learning based on classic semantic segmentation techniques. While significant progress has been made in breast cancer pathological image segmentation through the introduction of attention mechanisms, multi-scale feature modeling, and multi-task learning strategies, these approaches have not been universally successful.

[0003] However, for pathological images with complex tissue structures and significant scale variations, existing methods still have shortcomings in multi-scale feature representation and cross-scale information modeling, which can easily lead to the loss of feature information. At the same time, in areas with blurred tissue boundaries and easily confused categories, the lack of boundary guidance and cross-scale fusion mechanisms makes it difficult to maintain stability in terms of boundary continuity and detail representation of the segmentation results. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a multi-category tissue segmentation method for breast cancer pathological images, which solves the problems of insufficient multi-scale representation, blurred and complex boundaries between different types of tissues leading to unstable segmentation, and easy loss of detailed information during feature extraction and fusion in existing pathological image segmentation methods.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for multi-class tissue segmentation of breast cancer pathological images, comprising: Acquire pathological tissue images of the target patient; The pathological tissue image is input into a pre-trained multi-class tissue segmentation model for processing to obtain tissue segmentation results; the pre-training process of the multi-class tissue segmentation model includes: An image segmentation model is constructed that integrates an improved DenseASPP module and a boundary-aware dynamic weighted fusion mechanism; the improved DenseASPP module embeds a residual connection and a multi-scale adaptive pooling structure; the dynamic weighted fusion mechanism is embedded between the encoder and decoder of the image segmentation model; Build the training dataset; The image segmentation model is trained using the training dataset to obtain the multi-class tissue segmentation model.

[0006] Preferably, the improved DenseASPP module has a void ratio of .

[0007] Preferably, the processing procedure of the improved DenseASPP module includes: Feature blocks are constructed based on input features, and channel-dimensional dense connections are used to accumulate the feature blocks step by step to obtain multi-scale features; The multi-scale adaptive pooling structure is used to perform adaptive average pooling, channel projection, and upsampling back to the original resolution on the input features at multiple pooling scales to obtain context features. The multi-scale features and the context features are concatenated, projected, fused, and compressed into the output channel to obtain the main branch output. Extract the explicit residuals of the input features; The output characteristics are obtained by fusing the main branch output and the explicit residual.

[0008] Preferably, the expression for the multi-scale feature is: ;in, This refers to the multi-scale features described above; These are the multi-scale features described in the previous step; This represents a nonlinear transformation with an expansion rate d.

[0009] Preferably, the expression for the context feature is: ;in, For the context features; For upsampling; For convolutional projection; For adaptive average pooling; The input feature is referred to as "the input feature".

[0010] Preferably, the expression for obtaining the explicit residual is: ;in, For the explicit residual; The input feature; For batch normalization; for Convolution operation; Input the number of channels; This represents the number of output channels.

[0011] Preferably, the processing procedure of the dynamic weighted fusion mechanism includes: The received raw deep features and raw shallow features are projected onto a unified channel dimension to obtain aligned deep features and aligned shallow features; Boundary probability prediction is performed on the aligned shallow features to obtain a boundary probability map; Content-aware weights are predicted for the alignment deep features and the alignment shallow features to obtain initial fusion weights; A learnable scalar and a boundary gain are set, and the initial fusion weights are modulated using the learnable scalar and the boundary gain according to the boundary probability map to obtain modulated fusion weights; The modulation fusion weights are normalized, and the normalized modulation fusion weights are used to fuse the alignment deep features and alignment shallow features to obtain dynamic fusion features.

[0012] Preferably, the expression for the boundary probability map is: ;in, This is the boundary probability map; Use the Sigmoid activation function; Lightweight boundary detector; This refers to the alignment of shallow layer features.

[0013] Preferably, the expression for the initial fusion weights is: ;in, ; , These are the initial deep weights and the initial shallow weights, respectively. For normalization processing; This is the weighted prediction function; , These are the alignment deep features and the alignment shallow features, respectively.

[0014] Preferably, the expression for the modulation fusion weights includes: , ;in, ; , These are respectively modulated deep weights and modulated shallow weights; , These are the initial deep weights and the initial shallow weights, respectively. For element-wise multiplication; The boundary gain; It is a learnable scalar; This is the boundary probability map.

[0015] The present invention discloses the following technical effects: This invention provides a multi-class tissue segmentation method for breast cancer pathological images. By improving the DenseASPP module, it solves the feature loss problem caused by downsampling and dilated convolution in existing methods, and enhances the ability to model multi-scale and global contextual information, as well as improve the integrity and continuity of region prediction. Through a dynamic weighted fusion mechanism, it dynamically fuses features guided by boundary information, enhances boundary representation and the discriminative power at the junctions of various tissues, solves the problem of many boundary blurring and adhesion phenomena in existing methods, and improves the perception of boundary details and the semantic consistency of regions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the multi-category tissue segmentation process for breast cancer pathological images provided in an embodiment of the present invention; Figure 2 A schematic diagram of the RABF-SegNet network architecture and its improvement provided in the embodiments of the present invention; Figure 3 The DenseASPP architecture diagram is provided for the residual multi-scale adaptive pooling structure in the embodiments of the present invention. Figure 4 This is a diagram of the adaptive pooling AP module architecture provided in an embodiment of the present invention; Figure 5 This is a diagram of the BDWF module architecture provided in an embodiment of the present invention; Figure 6 A flowchart of boundary prediction provided for an embodiment of the present invention; Figure 7 This provides a visual comparison of ablation experiments of two innovative modules on a dataset, as provided in an embodiment of the present invention. Figure 8 The visualization results of the dataset provided in the embodiments of the present invention on the validation set. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The purpose of this invention is to provide a multi-category tissue segmentation method for breast cancer pathological images, which solves the problems of insufficient multi-scale representation, blurred and complex boundaries between different types of tissues leading to unstable segmentation, and easy loss of detailed information during feature extraction and fusion in existing pathological image segmentation methods.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Figure 1 This is a schematic diagram of the multi-category tissue segmentation process for breast cancer pathological images provided in an embodiment of the present invention, such as... Figure 1 As shown, the present invention provides a method for multi-class tissue segmentation of breast cancer pathological images, including: Step 100: Acquire pathological tissue images of the target patient; Step 200: Input the pathological tissue image into a pre-trained multi-class tissue segmentation model for processing to obtain tissue segmentation results; the pre-training process of the multi-class tissue segmentation model includes: Step 201: Construct an image segmentation model that integrates the improved DenseASPP module and the boundary-aware dynamic weighted fusion mechanism; the improved DenseASPP module embeds residual connections and multi-scale adaptive pooling structures; the dynamic weighted fusion mechanism is embedded between the encoder and decoder of the image segmentation model; Step 202: Construct the training dataset; Step 203: Train the image segmentation model using the training dataset to obtain the multi-class tissue segmentation model.

[0022] Specifically, the RABF-SegNet model architecture. The proposed model architecture diagram is shown below. Figure 2 As shown, this paper addresses the shortcomings of existing models in multi-scale feature extraction and limited boundary discrimination capabilities in two-dimensional medical pathology image segmentation tasks. Two specially designed modules are introduced at key locations to improve the segmentation performance of pathology images. The core innovation of RABF-SegNet lies in systematically solving two key bottlenecks in breast cancer image segmentation: 1) insufficient multi-scale feature representation for lesions with diverse morphologies; 2) blurred tissue boundaries and easy category confusion. Unlike existing methods that address these challenges, this embodiment combines two collaborative modules: RAP-DenseASPP for multi-scale feature capture and BDWF for enhancing boundary information representation and reducing category confusion, forming a unified framework to achieve superior segmentation performance in medical images.

[0023] Furthermore, unlike the original architecture, this embodiment inserts two innovative modules at key locations in the encoder and decoder: the RAP-DenseASPP module is embedded in... After the first layer, a five-branch parallel structure is used to achieve multi-scale feature extraction, including three densely connected dilated convolution branches (by...). Convolution and The system consists of dilated convolutions with dilation rates set to 1, 3, and 5, and two adaptive pooling (AP) branches, each employing... and Average pooling is applied. Simultaneously, residual connections are introduced to fuse input features with multi-scale feature representations; the BDWF module is embedded in... After the layer, it serves as a fusion bridge between deep and shallow layer features. This is achieved through... Convolution and The boundary detector, constructed using convolutions, generates a boundary probability map. By combining content-aware weights and boundary-aware weights, a dual weighting mechanism is built to achieve spatially adaptive feature fusion.

[0024] Preferably, the architecture design of this embodiment makes full use of the complementarity of deep and shallow features, which enhances the semantic expression capability while strengthening the fine-grained modeling capability of the boundary region.

[0025] refer to Figure 2 The network adopts an encoder-decoder framework, consisting of an encoder and a decoder. The encoder incorporates a RAP-DenseASPP module to enhance multi-scale semantic feature extraction capabilities; the decoder combines a content weight network with a boundary detector to achieve boundary-guided weighted fusion, highlighting boundary details and enhancing semantic information.

[0026] Specifically, this embodiment employs a residual multi-scale adaptive pooling (AP) structure called DenseASPP. Traditional ASPPs, when processing image segmentation of fine-grained targets and complex backgrounds, use a large hole ratio to expand the receptive field, leading to the loss of local detail features. To address this issue, this embodiment uses a DenseASPP module with a residual multi-scale pyramid pooling structure, as shown in the proposed module... Figure 3 As shown. By setting the void ratio to a more compact... First, by increasing local sampling density, over-smoothing is avoided, enhancing the capture of fine-grained textures and subtle boundaries. Second, a multi-scale adaptive average pooling (AP) module is introduced to enhance the semantic information of the global context while preserving local detail features. Finally, explicit residual connections are added to help preserve information about fine-grained structures and avoid the loss of details during multiple aggregation processes. This enables the DenseASPP module of residual multi-scale adaptive pooling (AP) to better handle the preservation of fine-grained structures, the capture of multi-scale features, and the modeling of boundary details in medical images, and to process images with complex backgrounds and subtle boundaries more efficiently.

[0027] Furthermore, given input features This embodiment uses a compact void ratio set. We construct feature blocks for multi-scale information extraction and achieve the stepwise accumulation of multi-scale features through channel-dimensional dense connections. : .

[0028] in, This represents a nonlinear transformation with an expansion rate d, which is derived from... Dimensionality reduction convolution and It is constructed by cascading dilated convolutions and obtained through normalization and activation functions.

[0029] Relying solely on dilated convolutions may lead to insufficient global context; therefore, this embodiment introduces a lightweight adaptive pooling AP module in parallel. For example... Figure 4 As shown, at multiple pooling scales The input features are adaptively pooled, channel-projected, and upsampled back to the original resolution. Finally, the context features are concatenated in the channel dimension and represented as follows: .

[0030] in, The output is Adaptive average pooling, for Convolution projection, This is for upsampling.

[0031] Then the terminal features of the DenseASPP main branch will be... With context features splicing, and through Projection layer fusion compression to The channel provides the main branch output. The output calculation is as follows: .

[0032] in, Used to integrate multi-scale representations and control channel size.

[0033] Since multi-scale aggregation may dilute the original discriminative cues in the input features, this embodiment introduces explicit residuals at the module output to stabilize gradient propagation and enhance the detail of the feature information. When using Convolution is performed with channel alignment; otherwise, an identity mapping is used. Final output. It can be represented as: .

[0034] in, For batch normalization; for Convolution operation.

[0035] By using explicit residual connections, the multi-scale information after complex transformations is ultimately preserved. and original input features The fusion of these features results in the following output characteristics: .

[0036] in, and These represent the number of input and output channels, respectively, referring to the channel dimension of data in operations such as convolutional layers and fully connected layers.

[0037] Preferably, a boundary-aware dynamic weighted fusion mechanism is used. In image segmentation of complex scenes, shallow and deep features have obvious complementarity. Shallow features usually retain richer spatial details and texture edge information, but they are difficult to capture sufficient global semantics and have limited expressive power; deep features have stronger semantic discriminative power, but their spatial resolution is lower and they are prone to losing details. Simply adding or splicing these features cannot maximize their advantages and may even lack the ability to model differences in spatial structure. This leads to the details in the boundary region being easily diluted by high-level semantic features, while the interior of the region may be affected by low-level noise. Therefore, the boundary region is not clearly represented, and the semantic consistency within the region is difficult to maintain.

[0038] Specifically, in order to overcome this limitation, the proposed BDWF module, such as Figure 5 As shown, it can predict the boundary probability and use it as a priori to spatially modulate the fusion weights, so that the boundary region focuses more on the expression of shallow details, while the internal region still maintains the semantically dominant fusion strategy. Figure 5 middle The normalized modulation weights.

[0039] Let the deep features be Shallow features are BDWF aims to achieve a boundary-friendly dynamic fusion output F in the spatial dimension. Firstly, this embodiment performs bilinear upsampling on deep features and then... Convolutional projection onto a uniform channel dimension C yields aligned features. and Represented as: .

[0040] .

[0041] in, All Convolutional projection. This alignment operation unifies the two feature representations in both spatial resolution and channel space, providing a consistent representational basis for subsequent pixel-by-pixel dynamic fusion.

[0042] Subsequently, considering that shallow features contain more significant local gradients and high-frequency texture information, such as Figure 6 As shown, this embodiment is based on aligned shallow features. Construct lightweight boundary branches to predict boundary probability graphs in a data-driven manner: .

[0043] in As a lightweight boundary detector, This is the Sigmoid activation function.

[0044] Based on this, this embodiment is based on the aligned features. Predict content-aware fusion weights. Specifically, concatenate the two feature streams along the channel dimension and then use a weight prediction function. Generate content weights corresponding to deep and shallow features. And normalization is performed in the channel dimension using softmax: .

[0045] .

[0046] The core of the BDWF module lies in spatially modulating content weights using boundary priors. To this end, this embodiment introduces a learnable scalar. Control the boundary enhancement intensity and define the boundary gain. At the boundary, some deep layer weights are transferred to shallow layer features, achieving a modulation representation as follows: .

[0047] .

[0048] in, This is for element-wise multiplication.

[0049] This weight modulation mechanism ensures that near boundary pixels, the fusion strategy favors preserving shallow details, while within the target region, it maintains a content-aware weighted fusion approach. To maintain numerical stability and ensure that the weights still constitute a pixel-by-pixel convex combination, this embodiment... and Normalization is then performed. Therefore, the dynamic fusion feature is represented as: .

[0050] Wherein, normalization is defined as , and The normalized modulation weights are used. The fused output is then subjected to BN normalization and ReLU activation to stabilize the feature distribution and introduce a nonlinear transformation.

[0051] By introducing a boundary-aware dynamic weighting mechanism, spatial differentiation of deep and shallow feature fusion behavior is achieved, guiding the boundary region to focus on shallow details, while emphasizing deep semantic information in the target's internal region, thus enabling feature fusion to form a position-adaptive fusion strategy in the spatial dimension.

[0052] Furthermore, evaluation metrics were used. This embodiment employed standard metrics such as average Dice coefficient (mDice), average intersection-union ratio (mIoU), average precision (mPrecision), and average recall (mRecall) to evaluate the method of this embodiment. The Dice coefficient for each category was also calculated to quantitatively evaluate the performance of semantic segmentation for each organization.

[0053] The Dice coefficient measures the degree of overlap between the predicted region and the actual region. The calculation formula is as follows: .

[0054] The average Dice coefficient represents the average degree of overlap for each category, reflecting the overall consistency of segmented regions.

[0055] .

[0056] The mean intersection-union ratio (mIoU) calculates the average of the intersection values ​​of the classes, and its expression is as follows: .

[0057] Average precision (mPre) represents the average precision for each category, expressed by the following formula: .

[0058] Average recall (Rec) is calculated as the average recall rate across all categories, and is expressed as follows: .

[0059] In the formula described above, C represents the total number of tissue categories. For any category c, this embodiment counts true positives at the pixel level ( ), false positives ( ) and false negatives ( TP represents the number of pixels that the model predicts as the target category and whose true labels also belong to that category, measuring the model's ability to identify and represent target regions. FP means the number of pixels that the model correctly identifies as not belonging to the target category, indicating the model's effective suppression of background or non-target regions. FN represents the number of pixels that the model predicts as not belonging to the target category, reflecting the model's missed detections and omissions of target regions.

[0060] Specifically, the experiments and results: The BCSS dataset, as it has been released, was collected from 151 patients diagnosed with breast cancer. The semantic segmentation of this dataset includes 151 whole-section pathological tissue images with H&E staining, each image containing semantic segmentation annotations for 20,000 tissue regions. Large breast cancer pathological sections were cropped to a pixel size of [missing information]. Image patches were generated by discarding pure background and filtering sub-images representing less than 1% of the dataset's categories to address class imbalance and reduce additional noise interference. The final processed data included categories such as Tumor (TUM), Stromal (STR), Lymphocytic Infiltration (LYM), Necrosis (NEC), and Other (OTR). A total of 100,000 pixel-level patches were generated. The dataset was randomly allocated into training, validation, and test sets, with the training set comprising 70%, the validation set 15%, and the remaining 15% automatically assigned to the test set.

[0061] Setup: All experiments were conducted on the same workstation running Windows 10. The GPU used was an NVIDIA GeForce RTX 4090 with 24GB of VRAM. The training framework used was PyTorch, with a pre-trained ResNet-50 network as the backbone. The AdamW optimizer was used, with a learning rate set to [missing value]. The weight decays to When training the BCSS dataset, the batch size was set to 8. Furthermore, to improve the model's generalization ability, data augmentation strategies were employed on the dataset: a random 90° rotation with a probability of 0.5, and horizontal and vertical flips with probabilities of 0.5 respectively; simultaneously, a deformation perturbation was randomly selected from ElasticTransform, GridDistortion, and OpticalDistortion with an overall probability of 0.3, and an illumination or noise enhancement was randomly selected from brightness / contrast, noise, and color with an overall probability of 0.3; finally, the input images were normalized.

[0062] Optionally, ablation studies are conducted. To verify the effectiveness of the proposed components in segmentation performance, this embodiment only changes the corresponding modules in the network structure, keeping the other hyperparameters consistent, and uses a unified metric to evaluate the model of this embodiment. This embodiment uses DeepLabV3+ as the baseline model, ResNet50 as the backbone network, and sequentially introduces the DenseASPP module (RAP-DenseASPP) with residual multi-scale adaptive pooling structure, and further adds a boundary-aware dynamic weighted fusion (BDWF) module to realize the RABF-SegNet model of this embodiment. This embodiment uses Dice coefficient, Intersection over Union (IoU), Precision, and Recall as metrics to evaluate the performance of the method. In addition, the Dice coefficient for each tissue category is calculated to analyze the model's segmentation ability for pathological tissue images.

[0063] First, these four methods were applied to the BCSS dataset for comparative analysis. For classic pathological images in this dataset, this embodiment provides qualitative visualization results, such as... Figure 7As shown in the visualization results, compared to the baseline model 1, model 2 enhances the multi-scale representation ability and global contextual information modeling ability of the model in multi-scale organizational structure regions, and reduces feature information loss. Model 3 exhibits clearer contours at organizational boundaries, while reducing missegmentation in organizational boundary regions. It can guide feature fusion through boundary-aware information, thereby improving boundary detail perception and regional semantic expression capabilities. Model 4, after fusing the two modules, performs better in terms of regional integrity, boundary clarity, and fine-grained structural representation. In the quantitative results of this embodiment, as shown in Tables 1 and 2, compared with the baseline model 1, models 2 and 3 improved the Dice coefficients on the BCSS dataset by 0.70% and 0.85%, respectively. The RABF-SegNet model of this embodiment improves the Dice, IoU, Precision, and Recall by 1.71%, 2.30%, 1.16%, and 1.62%, respectively, compared with the baseline model. In summary, the results show that the method of this embodiment can complement each other in terms of regional consistency and refined boundary segmentation, and can stably improve the performance of the model.

[0064] Table 1

[0065] Table 2

[0066] Furthermore, a comparative study was conducted. To verify the effectiveness of the method in this embodiment, this embodiment selected classic and latest breast cancer pathological image segmentation models such as U-Net, SwinTransformer, DeepLabV3, DeepLabV3+, BiSeNet, DETisSeg, and DCCUNet for comparative analysis. Experiments were carried out on five categories: tumor (TUM), stroma (STR), lymphocyte infiltration (LYM), inflammation (NEC), and other (OTR). Total Dice, IoU, Precision, and Recall were used as comprehensive evaluation indicators. The experimental results are shown in Table 3.

[0067] As shown in the table, the proposed RABF-SegNet outperforms state-of-the-art methods in all evaluation metrics. Quantitative evaluation results fully validate the significant performance of RABF-SegNet in multi-class segmentation of breast cancer tissue, providing more reliable technical support for the clinical auxiliary diagnosis of breast cancer. Among the five categories, the method in this embodiment achieved the best results with Dice values ​​of 0.9007, 0.6595, 0.8911, 0.8211, and 0.7948 for inflammation (NEC), other (OTR), tumor (TUM), stroma (STR), and lymphocyte infiltration (LYM), respectively. This indicates that the method in this embodiment has good generalization ability for different target categories, especially performing better in complex regions and easily confused categories. Overall, in this embodiment, the method achieved total Dice, IoU, Precision, and Recall of 0.8135, 0.6942, 0.8279, and 0.8034, respectively. Compared to the baseline model DeepLabV3+, the performance improvements were 1.71%, 2.30%, 1.16%, and 1.62%, respectively. Compared to the better-performing comparison model DCCUNet, the performance improvements were 1.29%, 1.79%, 1.54%, and 1.00%, respectively.

[0068] Table 3

[0069] Furthermore, this embodiment also selected several typical pathological images and visually compared the segmentation results of different models. From Figure 8 As can be seen, different models exhibit significant differences in target region boundary representation, fine-grained structure preservation, and category region integrity. Figure 8 The following is a visualization of the dataset on the validation set: Figure 8 (a) Partially corresponding input, Figure 8 (b) Partially corresponds to the real mask. Figure 8 (c) Part corresponds to DCCUNet, Figure 8 (d) corresponds to SwingTransformer. Figure 8 (e) corresponds to DETisSeg. Figure 8 (f) corresponds to DeepLabV3. Figure 8 (g) corresponds to BiseNet. Figure 8 (h) corresponds to DeepLabV3+. Figure 8 (i) Part corresponds to Unet, Figure 8 (j) corresponds to RABF-SegNet.

[0070] Specifically, while models such as Unet, DETisSeg, BiSeNet, Swin Transformer, and DCCUNet can identify the main target regions relatively well, they still exhibit varying degrees of missegmentation and omissions in areas with large scale variations, complex textures, or blurred class boundaries. Particularly in fine-grained tissue structures and edge transition regions, these models are prone to boundary blurring, missing local details, and class confusion, indicating their limited ability to represent complex pathological tissue morphologies. In contrast, DeepLabV3 and DeepLabV3+, due to their stronger dilated convolutional receptive field and multi-scale feature extraction capabilities, show improvements in region integrity and semantic consistency. However, they still suffer from local prediction discontinuities and insufficient boundary perception at elongated structures, small-area targets, and complex tissue boundaries, indicating deficiencies in global context supplementation and boundary detail modeling.

[0071] In the method proposed in this embodiment, the introduction of the DenseASPP module and BDWF module of residual multi-scale adaptive pooling enhances the ability of multi-scale feature information fusion, global context modeling, and boundary awareness, reducing the loss of detailed information, semantic aliasing, and region adhesion problems. This results in clearer and more continuous boundary segmentation and more accurate fine-grained structural representation. Furthermore, it can be observed that RABF-SegNet can still accurately distinguish adjacent category regions under complex background interference, exhibits better representation of fine structures and irregular boundaries, and demonstrates more stable prediction results and stronger regional consistency.

[0072] The beneficial effects of this invention are as follows: The proposed segmentation model, RABF-SegNet, enhances the representation of multi-scale tissue structures in breast cancer pathology images. While preserving global semantic information, it strengthens the expression of local boundary details, significantly alleviating problems such as boundary blurring, feature confusion, and region adhesion between different tissue categories in breast cancer pathology images, thereby achieving an overall improvement in multi-category tissue segmentation performance. This model mainly includes two core designs: First, it constructs a DenseASPP module that integrates residual connections and multi-scale adaptive pooling, employing a more compact dilatation rate to enhance the modeling capability of multi-scale and global contextual information, mitigating feature loss caused by downsampling and dilated convolution, and improving the completeness and continuity of region prediction; second, it uses a boundary-aware dynamic weighted fusion module (BDWF) to dynamically fuse feature information guided by boundary information, enhancing the discriminative power at the boundaries of different tissues, improving boundary detail awareness and regional semantic consistency, and reducing boundary blurring and adhesion. The proposed method was systematically evaluated using the Breast Cancer Semantic Segmentation dataset, validating its effectiveness. RABF-SegNet outperformed the baseline DeepLabV3+ by 1.71%, 2.30%, 1.16%, and 1.62% in total Dice, mIoU, mPrecision, and mRecall, respectively. Experimental results demonstrate that this method improves the segmentation accuracy of multi-class tissue regions in breast cancer pathological images while further emphasizing boundary representation capabilities, thus mitigating boundary blurring and adhesion between different tissue classes.

[0073] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0074] This embodiment uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for multi-class tissue segmentation in breast cancer pathological images, characterized in that, include: Acquire pathological tissue images of the target patient; The pathological tissue image is input into a pre-trained multi-class tissue segmentation model for processing to obtain the tissue segmentation result. The pre-training process of the multi-class tissue segmentation model includes: An image segmentation model is constructed that integrates an improved DenseASPP module and a boundary-aware dynamic weighted fusion mechanism; the improved DenseASPP module embeds a residual connection and a multi-scale adaptive pooling structure; the dynamic weighted fusion mechanism is embedded between the encoder and decoder of the image segmentation model; Build the training dataset; The image segmentation model is trained using the training dataset to obtain the multi-class tissue segmentation model.

2. The method for multi-class tissue segmentation of breast cancer pathological images according to claim 1, characterized in that, The improved DenseASPP module has a void ratio of .

3. The method for multi-class tissue segmentation of breast cancer pathological images according to claim 1, characterized in that, The processing procedure of the improved DenseASPP module includes: Feature blocks are constructed based on input features, and channel-dimensional dense connections are used to accumulate the feature blocks step by step to obtain multi-scale features; The multi-scale adaptive pooling structure is used to perform adaptive average pooling, channel projection, and upsampling back to the original resolution on the input features at multiple pooling scales to obtain context features. The multi-scale features and the context features are concatenated, projected, fused, and compressed into the output channel to obtain the main branch output. Extract the explicit residuals of the input features; The output characteristics are obtained by fusing the main branch output and the explicit residual.

4. The method for multi-class tissue segmentation of breast cancer pathological images according to claim 3, characterized in that, The expression for the multi-scale feature is: ;in, This refers to the multi-scale features described above; These are the multi-scale features described in the previous step; This represents a nonlinear transformation with an expansion rate d.

5. The method for multi-class tissue segmentation of breast cancer pathological images according to claim 3, characterized in that, The expression for the context feature is: ;in, For the context features; For upsampling; For convolutional projection; For adaptive average pooling; The input feature is referred to here.

6. The method for multi-class tissue segmentation of breast cancer pathological images according to claim 3, characterized in that, The expression for obtaining the explicit residual is: ;in, For the explicit residual; The input feature; For batch normalization; for Convolution operation; Input the number of channels; This represents the number of output channels.

7. The method for multi-class tissue segmentation of breast cancer pathological images according to claim 1, characterized in that, The processing steps of the dynamic weighted fusion mechanism include: The received raw deep features and raw shallow features are projected onto a unified channel dimension to obtain aligned deep features and aligned shallow features; Boundary probability prediction is performed on the aligned shallow features to obtain a boundary probability map; Content-aware weights are predicted for the alignment deep features and the alignment shallow features to obtain initial fusion weights; A learnable scalar and a boundary gain are set, and the initial fusion weights are modulated using the learnable scalar and the boundary gain according to the boundary probability map to obtain modulated fusion weights; The modulation fusion weights are normalized, and the normalized modulation fusion weights are used to fuse the alignment deep features and alignment shallow features to obtain dynamic fusion features.

8. The method for multi-category tissue segmentation of breast cancer pathological images according to claim 7, characterized in that, The expression for the boundary probability map is: ;in, This is the boundary probability map; Use the Sigmoid activation function; Lightweight boundary detector; This refers to the alignment of shallow layer features.

9. A method for multi-category tissue segmentation of breast cancer pathological images according to claim 7, characterized in that, The expression for the initial fusion weights is: ;in, ; , These are the initial deep weights and the initial shallow weights, respectively. For normalization processing; For weighted prediction functions; , These are the alignment deep features and the alignment shallow features, respectively.

10. A method for multi-class tissue segmentation of breast cancer pathological images according to claim 7, characterized in that, The expression for the modulation fusion weights includes: , ;in, ; , These are respectively modulated deep weights and modulated shallow weights; , These are the initial deep weights and the initial shallow weights, respectively. For element-wise multiplication; The boundary gain; It is a learnable scalar; This is the boundary probability map.

Images