A breast cancer image segmentation system based on double-track feature collaborative decoding

Abstract

The application discloses a breast cancer image segmentation system based on double-track feature cooperative decoding, which comprises a TransUNet encoding main body, a shallow feature screening enhancement module and a multistage decoder; the TransUNet is used as an encoder, and the encoder comprises an encoding stage and a multistage decoding stage, and the shallow feature screening enhancement module is arranged; a main body passage of the multistage decoding stage forms a main body decoding passage; the shallow feature screening enhancement module cooperates with a contour thinning module in each stage of the decoder to form an auxiliary thinning passage parallel to the main body decoding passage, and the two passages jointly form a double-track feature cooperative decoding structure; the model closes the semantic difference between the layers through a feature fusion enhancement module, captures lesion features of different scales through a multi-scale feature extraction module, and repairs an invasive fuzzy boundary through a contour thinning module; and the application can effectively filter background noise, enhance lesion response, improve the segmentation precision and robustness of a breast cancer lesion, and is suitable for clinical auxiliary diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image segmentation and computer vision technology, and in particular to a breast cancer image segmentation system based on dual-track feature collaborative decoding. Background Technology

[0002] Breast cancer is a highly prevalent malignant tumor in women. Early and accurate detection and boundary assessment are core prerequisites for reducing breast cancer mortality and developing individualized treatment plans. Early detection and precise segmentation of lesions through breast imaging (such as ultrasound, X-ray, or MRI) are of great significance for differentiating between benign and malignant breast cancer and for preoperative planning. Currently, relying on manual image review and lesion delineation suffers from low efficiency, high subjectivity, missed diagnoses due to fatigue, and large errors in boundary assessment. Therefore, deep learning-based automatic image segmentation algorithms have become the mainstream research direction for assisted diagnosis.

[0003] Current mainstream breast cancer segmentation networks are mostly based on U-Net and TransUNet architectures. TransUNet, in particular, combines the global modeling capabilities of the Transformer with the local detail extraction advantages of CNNs, exhibiting better performance in global lesion perception compared to traditional pure convolutional networks. However, the existing standard TransUNet still has significant shortcomings when used for breast cancer image segmentation:

[0004] (1) Background information such as glandular texture, fat gap and speckle noise in breast images will be transmitted layer by layer by the encoder and mixed into the features of each level, resulting in the effective features of small nodules or low-contrast lesions being submerged.

[0005] (2) There is a significant semantic gap between the shallow detail features and deep semantic features of the encoder. Simple feature combinations are not enough to fully model cross-level dependencies, resulting in inaccurate localization of small lesions and blurred boundaries of large masses.

[0006] (3) The size of breast cancer masses varies greatly, and the feature extraction of a fixed receptive field is difficult to cover multiple scale targets at the same time, which easily leads to missed detection of small lesions and incomplete segmentation of large tumors.

[0007] (4) Breast cancer masses often grow invasively and have severely degraded borders. Conventional decoding networks lack targeted border refinement mechanisms, resulting in rough and broken segmentation contours, which makes it difficult to meet the clinical need for accurate assessment of the extent of invasion.

[0008] The four types of problems mentioned above essentially stem from the lack of a multi-level collaborative optimization mechanism in the decoding stage: it is necessary to suppress tissue noise in encoder features, bridge cross-level semantic differences, perceive multi-scale lesions, and repair infiltrative blurred boundaries. Existing methods mostly rely on a single attention module or additional boundary loss for local repair, making it difficult to accomplish the above tasks simultaneously within a unified framework, resulting in an inability to simultaneously perceive global lesions and local contour details. Summary of the Invention

[0009] The purpose of this invention is to provide a breast cancer image segmentation system based on dual-track feature collaborative decoding, in order to solve the problems of low segmentation accuracy and insufficient robustness in existing breast cancer image segmentation technologies, which are caused by encoder features containing tissue noise, cross-level semantic gaps, large differences in lesion scale, and blurred invasive boundaries.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A breast cancer image segmentation system based on dual-track feature collaborative decoding includes a TransUNet encoding backbone, a shallow feature filtering and enhancement module, and a multi-level decoder. The multi-level decoder includes a feature fusion enhancement module, a multi-scale feature extraction module, and a contour refinement module. The TransUNet encoding backbone performs hierarchical encoding on the input image, outputting encoder features at four levels. The shallow feature filtering and enhancement module receives the shallow encoded features output by the encoder, generates shallow filtered and enhanced features, and then scale-adapts these features before sending them to the contour refinement modules in each level of the decoder. The shallow feature filtering and enhancement module adaptively filters encoded features to suppress background redundancy interference. The shallow feature filtering and enhancement module, in conjunction with the contour refinement module, forms an auxiliary refinement path parallel to the multi-level decoder backbone, constituting a dual-track feature collaborative decoding structure. The feature fusion enhancement module fuses cross-level features and strengthens feature semantic representation through a dual attention mechanism. The multi-scale feature extraction module constructs multiple receptive field extraction branches and completes cross-scale feature interaction. The contour refinement module focuses on the blurred boundary regions of breast cancer masses, supplementing edge details to optimize boundary segmentation.

[0012] In the system described above, the input to the shallow feature filtering and enhancement module is the shallowest level encoded feature and the second shallowest level encoded feature. The second shallowest level encoded feature is upsampled, and then the shallowest level encoded feature and the upsampled second shallowest level encoded feature are subjected to grouped convolution and batch normalization operations to align the spatial resolution and channel dimension of the two types of features. The aligned two types of features are concatenated along the channel dimension and then nonlinearly mapped using the ReLU activation function to obtain a fused feature. The fused feature is then sequentially processed through 1×1 pointwise convolution, batch normalization, and the Sigmoid function to generate an adaptive attention weight map. The attention weight map is then multiplied element-wise with the features output by the encoder at the corresponding level to output the shallow filtering and enhancement feature.

[0013] In the system described above, the input to the feature fusion enhancement module is the multi-level features output by the encoder; the multi-level features are denoted as the shallowest level encoded features in ascending order of depth. Second-shallow level coding features Second-deep level coding features and deepest level coding features The input to the first feature fusion enhancement module is... and The input to the second feature fusion enhancement module is , and The input to the third feature fusion enhancement module is , , and For each feature fusion enhancement module, the deepest encoded feature in its input is used as the baseline feature. The remaining encoded features are downsampled to the resolution of this baseline feature and then concatenated with the baseline feature along the channel dimension to obtain the initial fused feature. A 1×1 pointwise convolution is used to perform channel dimensionality reduction on the initial fused feature. The channel-dimensionality-reduced features are then input into the channel attention branch and the spatial attention branch in parallel. Channel attention branch processing: Global average pooling and global max pooling are performed on the channel-dimensionality-reduced features to extract channel description information. The two pooling results are added element-wise and then processed sequentially through a fully connected layer, a Swish activation function, a fully connected layer, and a Sigmoid function to generate a channel attention weight map. The channel attention weight map is then multiplied element-wise with the channel-dimensionality-reduced features to obtain the channel-enhanced features. Spatial attention branch processing: ... After channel dimensionality reduction, the features are subjected to mean pooling, max pooling, min pooling, and summation pooling, and multiple pooling results are concatenated to obtain pooled concatenated features. The pooled concatenated features are then subjected to 7×7 depthwise convolution, 1×1 pointwise convolution, Swish activation function, and Sigmoid function to obtain a spatial attention weight map. The spatial attention weight map is multiplied element-wise with the channel dimensionality reduction features to obtain spatial enhancement features. The channel enhancement features and spatial enhancement features are then subjected to bidirectional cross-enhancement operations: first, the spatial enhancement features are multiplied element-wise with the channel enhancement features to obtain the first interactive enhancement feature; then, the channel enhancement features are multiplied element-wise with the spatial enhancement features to obtain the second interactive enhancement feature. The two sets of interactive enhancement features are concatenated along the channel dimension, and then channel dimensionality reduction is completed through 1×1 pointwise convolution to output the fused enhancement features.

[0014] In the aforementioned system, within the multi-scale feature extraction module:

[0015] (1) The fused and enhanced features are input into multiple heterogeneous branches in parallel. The branches are in the following order: max pooling branch and multiple depthwise convolution branches with convolution kernel size arranged from small to large, which are used to capture the initial features under different receptive fields.

[0016] (2) Following the above branch order, multiply the initial features output by adjacent branches element by element to construct cross-scale interactive features and strengthen cross-scale information association;

[0017] (3) All cross-scale interactive features are spliced ​​along the channel dimension, and then the number of channels is adjusted by 1×1 pointwise convolution to output multi-scale fusion features.

[0018] In the system described above, the contour refinement module takes as input multi-scale fusion features, shallow filtering enhancement features, and segmentation prediction maps output by adjacent deep decoders. The segmentation prediction maps are upsampled to the current layer resolution, and a boundary blur weight map is calculated based on the upsampled prediction maps. The shallow filtering enhancement features output by the shallowest layer encoder are downsampled to the current layer resolution, and the boundary blur weight map is used to multiply the downsampled shallow filtering enhancement features output by the shallowest layer encoder element-wise to highlight the shallow details of the blurred boundary region of the breast cancer mass. The shallow filtering enhancement features output by the shallowest layer encoder after highlighting the shallow details are then added element-wise with the multi-scale fusion features, and finally, a 1×1 pointwise convolution is used to complete channel mapping, outputting the boundary enhancement features.

[0019] The system described above uses the CBIS-DDSM dataset for model training and performance testing. The dataset contains 2,620 pathologically confirmed mammograms with lesion areas annotated by professional radiologists. After preprocessing, the images are uniformly adjusted to a size of 256×256 pixels. The dataset is divided into training, validation, and test sets in a 7:1:2 ratio.

[0020] The system employs a joint loss function composed of a weighted coupling of focus loss and Dice loss during the model training phase, with the optimization objective of minimizing the loss value. All learnable parameters of the network are iteratively updated through the backpropagation algorithm. The model segmentation performance is quantitatively evaluated using three indicators: Dice similarity coefficient, average intersection-union ratio, and pixel accuracy. During training, a validation set is partitioned to monitor the model's generalization performance in real time, and a learning rate decay strategy is used to suppress overfitting.

[0021] A breast cancer image segmentation method based on dual-track feature collaborative decoding uses TransUNet as the encoder to extract features from the input breast cancer image and perform progressive downsampling, outputting multi-level encoded features. The network consists of an encoding stage, a shallow feature selection and enhancement module, and a multi-level decoding stage. The multi-level decoding stage forms the main decoding path based on multi-level encoded features, while the shallow feature selection and enhancement module works in conjunction with contour refinement modules at each level to form an auxiliary refinement path. Together, they constitute a dual-track feature collaborative decoding structure.

[0022] The shallow feature filtering and enhancement module receives shallow encoded features, generates an adaptive attention weight map, and performs pixel-level weighting on the shallow encoded features. It filters out background noise and enhances lesion response, outputting shallow filtered and enhanced features. After scale adaptation, these features are sent to contour refinement modules at various levels to provide clean shallow details for boundary repair.

[0023] The feature fusion enhancement module progressively fuses the multi-level encoded features output by the encoder during the multi-level decoding stage. This module uses the deepest encoded feature as a baseline, aligns the remaining encoded features, concatenates them for dimensionality reduction, and then bridges cross-level semantic differences through a dual mechanism of channel attention and spatial attention, as well as bidirectional cross-enhancement operations, outputting fused and enhanced features.

[0024] The multi-scale feature extraction module feeds the input features into multiple heterogeneous branches in parallel to capture features under different receptive fields, and constructs multi-scale fusion features through a cross-scale interaction mechanism to take into account the feature extraction of lesions of different sizes and reduce the risk of missed detection of small lesions.

[0025] The contour refinement module receives multi-scale fused features, scale-adapted shallow-layer filtered and enhanced features, and segmentation predictions passed from adjacent deep-layer decoding units at each level of the decoding unit. Figure 3 The input is processed by a deep semantic layer to guide the supplementation of shallow details, collaboratively repairing boundaries and generating intermediate segmentation prediction maps. The segmentation prediction map output by the deepest level decoding unit is upsampled to restore the original image resolution, resulting in the segmentation result of the breast cancer lesion.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] A dual-track feature collaborative decoding structure is adopted. The main decoding path relies only on multi-level encoded features, while the auxiliary refinement path introduces shallow filtering enhancement features and deep prediction maps to avoid interference from irrelevant information caused by traditional skip connections.

[0028] The shallow feature screening enhancement module filters out background noise at the source, enhances lesion response, and improves the effectiveness of shallow features;

[0029] The feature fusion enhancement module effectively bridges cross-level semantic differences and improves feature discriminative power through a dual attention mechanism and bidirectional cross-enhancement operation;

[0030] The multi-scale feature extraction module takes into account the feature extraction of lesions of different sizes through multi-path heterogeneous branches and cross-scale interaction mechanism, reducing the risk of missed detection of small lesions;

[0031] The contour refinement module integrates multi-scale features, shallow enhancement features, and deep prediction maps. It uses deep semantics to guide the supplementation of shallow details, effectively repairing the problem of blurred infiltrative boundaries and improving segmentation accuracy and robustness. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the dual-track feature collaborative decoding network proposed in this invention;

[0033] Figure 2 This is a schematic diagram of the shallow feature filtering enhancement module in this invention;

[0034] Figure 3 This is a schematic diagram of the feature fusion enhancement module in this invention;

[0035] Figure 4 This is a schematic diagram of the multi-scale feature extraction module in this invention;

[0036] Figure 5 This is a schematic diagram of the contour refinement module in this invention. Detailed Implementation

[0037] The present invention will be described in detail below with reference to specific embodiments.

[0038] This invention provides a breast cancer image segmentation system based on dual-track feature collaborative decoding. A breast cancer image segmentation model is constructed and used to complete breast cancer image segmentation. The model uses TransUNet as the encoder, including an encoding stage and a multi-level decoding stage, and incorporates a shallow feature filtering enhancement module. The backbone path of the multi-level decoding stage constitutes the backbone decoding path; the shallow feature filtering enhancement module cooperates with the contour refinement modules in each level of the decoder to form an auxiliary refinement path parallel to the backbone decoding path; the backbone decoding path and the auxiliary refinement path together constitute a dual-track feature collaborative decoding structure.

[0039] In the encoding stage, the input breast cancer image is downsampled layer by layer and features are extracted, outputting four levels of encoded features, namely the shallowest level encoded features. Second-shallow level coding features Second-deep level coding features and deepest level coding features .

[0040] This invention includes an independent shallow feature filtering enhancement module, which receives... and First of all, Perform upsampling to make its spatial resolution the same as the previous one. Alignment. Then, 3×3 grouped convolution and batch normalization are performed on both types of features respectively to expand the spatial receptive field and control the number of parameters, while unifying the number of channels. The processed two features are then concatenated along the channel dimension, and a non-linear mapping is completed using the ReLU activation function to obtain the fused feature, the formula of which is shown below:

[0041] (1)

[0042] In formula (1), g and x represent the values ​​after upsampling, respectively. and , and This represents a 3×3 grouped convolution, BN represents batch normalization, Concat represents concatenation along the channel dimension, and ReLU represents the activation function. This indicates the fusion feature.

[0043] Next, After sequential 1×1 pointwise convolution, batch normalization, and Sigmoid function operations, an adaptive attention weight map with values ​​ranging from 0 to 1 is generated. This attention weight map is then multiplied element-wise with the features output by the corresponding level encoder to output shallow-layer filtering and enhancement features, as shown in the following formula:

[0044] (2)

[0045] In formula (2), This indicates shallow screening to enhance features. This indicates an element-wise multiplication operation. This represents a 1×1 pointwise convolution. This represents the Sigmoid function.

[0046] In the multi-level decoding stage, the decoder backbone is sequentially configured with a feature fusion enhancement module and a multi-scale feature extraction module; except for the deepest level decoding stage, each other decoding level is configured with a contour refinement module; the deepest level decoder does not have a contour refinement module. The contour refinement module within each level decoder receives multi-scale fused features from the backbone path, segmentation prediction maps output from adjacent deep decoders, and shallow-level filtered and enhanced features from the aforementioned shallow feature filtering and enhancement module after scale adaptation.

[0047] In the multi-level decoding stage, the feature fusion enhancement modules corresponding to different levels of decoders receive different combinations of multi-level encoded features. The input to the first feature fusion enhancement module is... and The input to the second feature fusion enhancement module is , and The input to the third feature fusion enhancement module is , , and .

[0048] The design goal of the feature fusion enhancement module is to selectively enhance high-value feature channels and high-information features in the spatial dimension, while aggregating encoded features from corresponding levels to reduce the semantic gap between different levels. Taking a four-level decoding structure as an example, level 1 is the shallowest level, and level 4 is the deepest level. Level i ( The feature fusion enhancement module receives the encoded features of the corresponding layer and also aggregates the features output by all preceding layer decoders. The specific work process is as follows:

[0049] First, downsampling to After concatenating the features along the channel dimension at the same resolution, the initial fused features are obtained. A 1×1 pointwise convolution is then used to perform channel-wise dimensionality reduction on the initial fused features, resulting in the channel-wise reduced features, as shown in the formula below:

[0050] (3)

[0051] In formula (3), when i=2, only the output features of the first level are aggregated. , This represents the features of the channel after dimensionality reduction.

[0052] In the channel attention branch, firstly... Global average pooling and global max pooling are performed separately to extract the overall contextual information and key activation information of the features. After element-wise addition of the two pooling results, the mixture is processed sequentially through a fully connected layer, a Swish activation function, and another fully connected layer. The Swish activation function introduces non-linearity and mitigates gradient abrupt changes. Subsequently, a channel attention weight map is generated using the Sigmoid function. This channel attention weight map is then compared with... Element-wise multiplication yields the channel enhancement features, with formulas shown in (4) and (5) below:

[0053] (4)

[0054] (5)

[0055] In formula (4), FC represents the fully connected layer, and GAP and GMP represent global average pooling and global max pooling, respectively. Represents the channel attention weight map;

[0056] In formula (5), This indicates channel enhancement features.

[0057] Meanwhile, in spatial attention branch processing, for Mean pooling, max pooling, min pooling, and summation pooling are performed along the channel dimension to capture the existence, significance, weakest response, and overall strength information of spatial features, respectively. The pooled outputs are then concatenated to obtain the pooled concatenated features, as shown in the following formula:

[0058] (6)

[0059] In formula (6), , , , These represent mean pooling, max pooling, min pooling, and summation pooling operations along the channel dimension, respectively. This represents the pooling splicing feature.

[0060] The spatial attention weight map is obtained by sequentially processing the data through 7×7 depthwise convolution, 1×1 pointwise convolution, Swish activation function, and Sigmoid function. This spatial attention weight map is then compared with... Element-wise multiplication yields the spatially enhanced features, as shown in formulas (7) and (8) below:

[0061] (7)

[0062] (8)

[0063] In formula (7), Represents the spatial attention weight map. This represents a 7×7 depthwise convolution;

[0064] In formula (8), This represents spatially enhanced features.

[0065] Subsequently, and Entering the two-way cross-interaction mechanism. In the first interaction path, and Element-wise multiplication achieves cross-dimensional feature interaction with channel features as the main component and spatial features as the modulator, resulting in the first interactive enhancement feature; in the second interactive path, and Element-wise multiplication is performed to achieve cross-dimensional feature interaction with spatial features as the main component and channel features as the modulator, resulting in the second interactive enhancement feature. The formulas are shown in (9) and (10) below:

[0066] (9)

[0067] (10)

[0068] In formula (9), C represents the first interactive enhancement feature;

[0069] In formula (10), D represents the second interactive enhancement feature.

[0070] Concatenate C and D along the channel dimension, then compress the number of channels using a 1×1 pointwise convolution to output the fused enhancement feature of the i-th level decoder, as shown in the formula below:

[0071] (11)

[0072] In formula (11), This represents the fusion enhancement feature of the i-th level decoder.

[0073] The network then proceeds to the multi-scale feature extraction module. This module extracts features in parallel through multi-scale receptive fields and establishes strong cross-scale coupling relationships to enhance the network's ability to extract local detailed features and alleviate the problem of losing small breast lesion features during encoder downsampling. For the shallowest level decoding stage, since there are no preceding shallow features to aggregate, the feature fusion enhancement module is omitted in the decoder backbone. Only the multi-scale feature extraction module and contour refinement module are configured to directly perform multi-scale parsing and boundary repair on the input features of this level.

[0074] Specifically, for layers configured with feature fusion enhancement modules, Four extraction branches are input in parallel: max pooling branch, 3×3 depthwise convolution branch, 5×5 depthwise convolution branch, and 7×7 depthwise convolution branch, which extract multi-scale features corresponding to different receptive fields. Among them, the max pooling branch is used to enhance the local texture information in the features, and the depthwise convolution branches of different sizes capture fine-grained local details, medium-range context, and large-range spatial dependency information, respectively. For the shallowest level decoding stage, since the feature fusion enhancement module is omitted, the input features of this level are directly input into the above extraction branches in parallel.

[0075] To establish strong coupling between cross-scale features, the initial features output from every two adjacent level branches are multiplied element-wise to achieve hierarchical information interaction between features at each level, resulting in cross-scale interactive features. The formulas are shown in (12), (13), (14) and (15) respectively:

[0076] (12)

[0077] (13)

[0078] (14)

[0079] (15)

[0080] In formulas (12) to (15), , , , Indicates cross-scale interaction features. , , , Indicates the initial characteristics.

[0081] All cross-scale interactive features are concatenated along the channel dimension, and then cross-channel information fusion is achieved through 1×1 pointwise convolution to reduce the number of channels, outputting multi-scale fused features, the formula of which is shown below:

[0082] (16)

[0083] In formula (16), This represents the multi-scale fusion feature of the i-th level decoder.

[0084] As the output feature of the current i-th level decoder This is used by subsequent, deeper-level feature fusion and enhancement modules for aggregation.

[0085] To address the challenge of blurred boundaries in breast cancer tumors, a contour refinement module is designed. This module compensates for blurred areas using shallow-level features and re-extracts features, enhancing the network's segmentation accuracy of breast cancer tumor contour boundaries. The contour refinement module is only applied to decoders at all levels except the deepest level; the deepest level decoder does not perform contour refinement because it lacks prediction images from subsequent levels as guiding information.

[0086] Specifically, for the i-th level decoder, the boundary blur weights are first calculated using the prediction map output by the next level decoder (upsampled to the current level resolution), as shown in the following formula:

[0087] (17)

[0088] In formula (17), This represents the boundary fuzzy weight of the i-th level; This represents the predicted map output by the (i+1)th level decoder; regions with predicted values ​​close to 0.5 are blurred regions with weights close to 1; regions with predicted values ​​close to 0 or 1 are defined background or foreground regions with weights close to 0.

[0089] Subsequently, a downsampling operation is performed on the shallow-level filtered enhancement features output by the shallowest-level encoder to ensure that their resolution is consistent with the current-level features; As guiding information, the shallow filtering enhancement features output by the shallowest level encoder, aligned with the resolution, are multiplied element-wise to extract detailed features from blurred boundary regions. These features are then added element-wise to the multi-scale fusion features of the current level to fuse shallow-level detailed information with deep-level semantic information. Finally, a 1×1 pointwise convolution is used to achieve cross-channel information fusion and output boundary enhancement features, as shown in the following formula:

[0090] (18)

[0091] In formula (18), This represents the boundary enhancement feature of the i-th level. This represents the shallowest level of shallow filtering enhancement features.

[0092] For the deepest level decoder, the multi-scale fused features output by the multi-scale feature extraction module are directly used as the final output of this level decoder to generate the deepest level prediction map.

[0093] For decoders at all levels except the deepest level, boundary enhancement features As the final output of the decoder at this level, it is used to generate the prediction map of the current level and pass the prediction map of the current level to the adjacent shallow level decoder.

[0094] During model training, a joint loss function consisting of a weighted coupling of focus loss and Dice loss is used as the optimization objective. The focus loss, by introducing a dynamic scaling factor into the standard cross-entropy loss, reduces the weight of simple negative samples, allowing the model to focus more on difficult-to-classify samples such as small lesions and regions with blurred boundaries. The Dice loss directly measures the overlap between the predicted segmented region and the gold standard annotation, enhancing the fit of the overall lesion contour. The calculation formulas for both are shown in (19) and (20), respectively:

[0095] (19)

[0096] (20)

[0097] In formulas (19) and (20), N represents the total number of pixels in a single image. Let i be the true label value of the i-th pixel. Predict the probability value for the i-th pixel using the model; This is a category balancing factor used to adjust the weight ratio of positive and negative samples; To focus on parameters, the weighting of easily classified samples is reduced, while the optimization contribution of difficult-to-classify samples is amplified.

[0098] With the goal of minimizing this joint loss, the model updates all learnable parameters round by round on the training set using the backpropagation algorithm. During training, a validation set is introduced simultaneously to monitor the model's generalization state, and hyperparameters such as the learning rate are dynamically adjusted to suppress overfitting.

[0099] The formula for calculating the joint loss function is shown below:

[0100] (twenty one)

[0101] In formula (21), , This represents the balance weighting coefficient between the two types of losses, used to adjust the optimization emphasis of the two losses.

[0102] To comprehensively evaluate the overall performance of the proposed model in medical image segmentation tasks, three general segmentation evaluation metrics were selected for quantitative evaluation: Dice similarity coefficient, average intersection-union ratio, and pixel accuracy. The three metrics reflect the model's segmentation effect from three dimensions: region overlap, global segmentation accuracy, and pixel classification accuracy, respectively. Their specific calculation formulas are shown in (22), (23), and (24) below.

[0103] (twenty two)

[0104] (twenty three)

[0105] In formulas (22) and (23), TP, FP, and FN correspond to the number of true positive pixels, the number of false positive pixels, and the number of false negative pixels, respectively.

[0106] (twenty four)

[0107] In formula (24), TN represents the number of true negative pixels.

[0108] The CBIS-DDSM dataset was divided into training, validation, and test sets according to a preset ratio of 7:1:2 and then input into the network to complete training and testing. The comparative test results of the method proposed in this embodiment are shown in Table 1.

[0109]

[0110] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A breast cancer image segmentation system based on dual-track feature collaborative decoding, characterized in that, The system comprises a TransUNet encoding backbone, a shallow feature filtering and enhancement module, and a multi-level decoder. The multi-level decoder includes a feature fusion enhancement module, a multi-scale feature extraction module, and a contour refinement module. The TransUNet encoding backbone performs hierarchical encoding on the input image, outputting encoder features at four levels. The shallow feature filtering and enhancement module receives the shallow encoded features output by the encoder, generates shallow filtered and enhanced features, and then scale-adapts these features before sending them to the contour refinement modules in each level of the decoder. The shallow feature filtering and enhancement module adaptively filters encoded features and suppresses background redundancy interference. The shallow feature filtering and enhancement module, in conjunction with the contour refinement module, forms an auxiliary refinement path parallel to the multi-level decoder backbone, constituting a dual-track feature collaborative decoding structure. The feature fusion enhancement module fuses cross-level features and strengthens feature semantic representation through a dual attention mechanism. The multi-scale feature extraction module constructs multiple receptive field extraction branches and completes cross-scale feature interaction. The contour refinement module focuses on the blurred boundary region of breast cancer masses, supplementing edge details to optimize boundary segmentation.

2. The system according to claim 1, characterized in that, The input to the shallow feature filtering and enhancement module is the shallowest level coding feature and the second shallowest level coding feature. Upsampling is performed on the second shallowest level coding feature, and then group convolution and batch normalization are performed on both the shallowest level coding feature and the upsampled second shallowest level coding feature to align the spatial resolution and channel dimension of the two types of features. The aligned two types of features are then concatenated along the channel dimension, and a nonlinear mapping is performed using the ReLU activation function to obtain the fused feature. The fused features are sequentially processed through 1×1 pointwise convolution, batch normalization, and Sigmoid function operations to generate an adaptive attention weight map. The attention weight map is then multiplied element-wise with the features output by the encoder at the corresponding level to output shallow filtering and enhancement features.

3. The system according to claim 1, characterized in that, The input to the feature fusion enhancement module is the multi-level features output by the encoder; the multi-level features are denoted as the shallowest level encoded features in order of increasing depth. Second-shallow level coding features Second-deep level coding features and deepest level coding features The input to the first feature fusion enhancement module is... and The input to the second feature fusion enhancement module is , and The input to the third feature fusion enhancement module is , , and For each feature fusion enhancement module, the deepest encoded feature in its input is used as the baseline feature. The remaining encoded features are downsampled to the resolution of this baseline feature and then concatenated with the baseline feature along the channel dimension to obtain the initial fused feature. A 1×1 pointwise convolution is used to perform channel dimensionality reduction on the initial fused feature. The channel-dimensionality-reduced features are then input into the channel attention branch and the spatial attention branch in parallel. Channel attention branch processing: Global average pooling and global max pooling are performed on the channel-dimensionality-reduced features to extract channel description information. The two pooling results are added element-wise and then processed sequentially through a fully connected layer, a Swish activation function, a fully connected layer, and a Sigmoid function to generate a channel attention weight map. The channel attention weight map is then multiplied element-wise with the channel-dimensionality-reduced features to obtain the channel-enhanced features. Spatial attention branch processing: ... After channel dimensionality reduction, the features are subjected to mean pooling, max pooling, min pooling, and summation pooling, and multiple pooling results are concatenated to obtain pooled concatenated features. The pooled concatenated features are then subjected to 7×7 depthwise convolution, 1×1 pointwise convolution, Swish activation function, and Sigmoid function to obtain a spatial attention weight map. The spatial attention weight map is multiplied element-wise with the channel dimensionality reduction features to obtain spatial enhancement features. The channel enhancement features and spatial enhancement features are then subjected to bidirectional cross-enhancement operations: first, the spatial enhancement features are multiplied element-wise with the channel enhancement features to obtain the first interactive enhancement feature; then, the channel enhancement features are multiplied element-wise with the spatial enhancement features to obtain the second interactive enhancement feature. The two sets of interactive enhancement features are concatenated along the channel dimension, and then channel dimensionality reduction is completed through 1×1 pointwise convolution to output the fused enhancement features.

4. The system according to claim 1, characterized in that, In the multi-scale feature extraction module: (1) the fused and enhanced features are input into multiple heterogeneous branches in parallel. The branches are in the following order: max pooling branch and multiple depthwise convolution branches with convolution kernel size arranged from small to large, which are used to capture the initial features under different receptive fields. (2) Following the above branch order, multiply the initial features output by adjacent branches element by element to construct cross-scale interactive features and strengthen cross-scale information association; (3) Concatenate all cross-scale interactive features along the channel dimension, and then adjust the number of channels through 1×1 pointwise convolution to output multi-scale fusion features.

5. The system according to claim 1, characterized in that, The contour refinement module takes as input multi-scale fusion features, shallow filtering enhancement features, and segmentation prediction maps output by adjacent deep decoders. The segmentation prediction maps are upsampled to the current layer resolution, and a boundary blur weight map is calculated based on the upsampled prediction maps. The shallow filtering enhancement features output by the shallowest layer encoder are downsampled to the current layer resolution, and the boundary blur weight map is used to multiply the downsampled shallow filtering enhancement features output by the shallowest layer encoder element-wise to highlight shallow details in the blurred boundary region of the breast cancer mass. The shallow filtering enhancement features output by the shallowest layer encoder after highlighting shallow details are then added element-wise with the multi-scale fusion features, and finally, a 1×1 pointwise convolution is performed to complete channel mapping, outputting boundary enhancement features.

6. The system according to claim 1, characterized in that, Model training and performance testing used the CBIS-DDSM dataset, which contains 2,620 pathologically confirmed mammograms with lesion areas annotated by professional radiologists. After preprocessing, the images were uniformly adjusted to a size of 256×256 pixels. The dataset was divided into training, validation, and test sets in a 7:1:2 ratio.

7. The system according to claim 1, characterized in that, During the model training phase, a joint loss function is constructed by weighted coupling of focus loss and Dice loss, with the optimization objective being to minimize the loss value. All learnable parameters of the network are iteratively updated through the backpropagation algorithm. The model segmentation performance is quantitatively evaluated using three metrics: Dice similarity coefficient, average intersection-union ratio, and pixel accuracy. During training, a validation set is defined to monitor the model's generalization performance in real time, and a learning rate decay strategy is used to suppress overfitting.

Images