A three-dimensional breast tumor segmentation method based on cross-view consistency

Abstract

The application belongs to the technical field of image segmentation, and relates to a three-dimensional breast tumor segmentation method based on cross-view consistency. Firstly, dynamic contrast-enhanced magnetic resonance three-dimensional voxel data and a real segmentation mask are acquired, two-dimensional sequences and corresponding segmentation masks are obtained by slicing along axial, sagittal and coronal planes, local context splicing is performed on each view, enhanced features are extracted and then input into a segmentation network sharing weights, two-dimensional probability maps are output, segmentation loss and spatial smoothing constraint loss are calculated, a training model is jointly optimized, when testing, each view probability map is aligned to a three-dimensional coordinate system, and after average fusion and binarization, a three-dimensional breast tumor segmentation result is obtained. The algorithm solves the problems of segmentation collapse caused by insufficient training data under few samples and three-dimensional model fault and smoothness loss caused by two-dimensional multi-view receptive field limitation.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation technology, and relates to a cross-view... Figure 1 A three-dimensional segmentation method for homogeneous breast tumors. Background Technology

[0002] Dynamic contrast-enhanced magnetic resonance imaging (MRI) is considered the most sensitive imaging modality for detecting breast cancer, allowing for precise assessment of tumor hemodynamic characteristics. Providing accurate breast tumor annotation at the pixel level is crucial for providing essential information such as tumor localization, contour, and texture to support clinical diagnosis and computer-aided systems. Recently, with the rapid development of deep learning, 3D convolutional neural networks and Transformers have shown great potential in medical image segmentation. However, these massive 3D architectures inevitably require huge amounts of high-quality medical image data to achieve satisfactory performance and generalization capabilities, which severely conflicts with the scarcity of high-quality medical data in real-world clinical settings.

[0003] To address the aforementioned data scarcity dilemma, existing technical solutions mainly fall into two categories: one relies on self-supervised learning, artificially expanding the training set by generating pseudo-labels for unlabeled data; the other employs multi-view learning, dividing 3D medical images into multiple 2D slices along different views for independent processing. However, when these conventional solutions attempt to be implemented in actual intelligent medical auxiliary diagnostic systems, two insurmountable technical problems arise: (1) the segmentation collapse problem under extremely limited sample conditions. Pseudo-label methods heavily rely on generation quality. Under extremely limited clinical samples, low-quality pseudo-labels often severely mislead the optimization process, causing the model performance to not only fail to improve but also decline sharply; (2) the problem of discontinuity and lack of smoothness in 3D reconstruction (lack of 3D smoothness). Although existing multi-view methods effectively increase the sample size, they are limited by the inherent receptive field of 2D convolution kernels, and the segmentation model cannot effectively associate orthogonal contextual information between slices. This results in a rough surface and severe geometric discontinuity between slices in the final generated 3D tumor model, which cannot be directly applied to high-precision surgical navigation or target delineation.

[0004] To address the aforementioned problems, this invention proposes a method based on cross-view... Figure 1 A three-dimensional segmentation method for homogeneous breast tumors. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-view-based... Figure 1 A consistent three-dimensional breast tumor segmentation method is proposed to address the technical problems faced by existing medical assistance systems in practical clinical applications, such as low segmentation accuracy with very few samples and the lack of tortuosity and smoothness in three-dimensional reconstruction.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a cross-view... Figure 1 A three-dimensional segmentation method for homogeneous breast tumors includes the following steps: S1: Acquiring dynamic contrast-enhanced magnetic resonance three-dimensional voxel sample data and the corresponding real segmentation mask Y Slicing operations are performed along three mutually perpendicular orthogonal views to obtain two-dimensional slice sequences and corresponding two-dimensional real segmentation mask sequences under axial, sagittal, and coronal views, respectively. S2: Perform local context-aware stitching on the two-dimensional slice sequences under the three views respectively to obtain the enhanced slice features under each view; S3: Construct a convolutional neural network segmentation model and initialize its parameters. The convolutional neural network segmentation model shares the same weights among the three views. The enhanced slice features under each view are input into the convolutional neural network segmentation model to obtain the two-dimensional segmentation probability map under the corresponding view. S4: Calculate the segmentation loss Spatial similarity constraint loss and multiple views Figure 1 Consistency constraint loss The objective loss function is jointly optimized, and the parameters of the convolutional neural network segmentation model are updated using the jointly optimized objective loss function to obtain the trained convolutional neural network segmentation model. S5: Input the test sample into the trained convolutional neural network segmentation model to obtain the three-dimensional breast tumor segmentation result.

[0007] Furthermore, the specific steps of S1 are as follows: S1.1: Acquiring dynamic contrast-enhanced magnetic resonance three-dimensional voxel sample data and the corresponding real segmentation mask Y ; S1.2: Transfer the three-dimensional voxel sample data Slice along three mutually perpendicular orthogonal views to generate a 2D slice sequence in the three views, and then apply the result to the actual segmentation mask. Y Perform a slicing operation to generate a sequence of two-dimensional real segmentation masks in three views.

[0008] Furthermore, the specific steps of S2 are as follows: S2.1: For each view, traverse each slice in its two-dimensional slice sequence, define the current slice as the target slice, and extract its adjacent upper and lower slices; S2.2: The target slice is stitched and padded with its adjacent slices above and below it in the channel dimension to obtain the stitched multi-channel data; S2.3: Construct a residual convolution module, input the spliced ​​multi-channel data into the residual convolution module for feature extraction, and obtain the enhanced slice features under the three views respectively.

[0009] Furthermore, the specific steps of S3 are as follows: S3.1: Construct a convolutional neural network segmentation model based on the U-Net architecture; S3.2: Input the enhanced slice features from the three views into the convolutional neural network segmentation model to obtain two-dimensional segmentation probability maps from the three views respectively.

[0010] Furthermore, the specific steps of S4 are as follows: S4.1: Define the segmentation loss As shown in Equation (1), the consistency between the two-dimensional segmentation probability map output by the model and the two-dimensional true segmentation mask under the corresponding view is constrained, and the segmentation loss is... It is composed of Dice loss and mean square error loss, as shown in Equation (2) and Equation (3) respectively; (1) in: To divide the loss, For Dice's loss, For mean square error loss, This is a two-dimensional segmentation probability map predicted by the model. A two-dimensional true segmentation mask slice at the same location in the same view, corresponding to the two-dimensional segmentation probability map predicted by the model; (2) in: This represents the total number of pixels in a single slice. For pixel index, For the first The predicted probability of each pixel. For the first A true segmentation mask of 1 pixel, It is a smoothing factor; (3) S4.2: Define spatial similarity constraint loss As shown in equation (4); (4) in: This represents the total number of slices for the sample in this view. This is the probability map output by the convolutional neural network segmentation model. For the first Features of each slice For the first Features of each slice To control the Gaussian hyperparameter of the distance weight decay rate, As the normalization factor, For the slice sequence A slice, For the slice sequence m One slice; S4.3: Defining Multiple Views Figure 1 Consistency constraint loss As shown in equation (5), it is used to prevent strong smoothing constraints from causing the model to output a trivial solution where all slices are the same. (5) in: For multiple views Figure 1 Consistency constraint loss, For a collection of view pairs, Let be the coordinates of any voxel in three-dimensional space. ) for view In three-dimensional coordinates The predicted probability of the mapping at the location, Represents a three-dimensional spatial domain. The total number of voxels. For in view In three-dimensional coordinates The corresponding predicted probability value, The square of the Euclidean distance; S4.4: The segmentation loss Spatial similarity constraint loss and multiple views Figure 1 Consistency constraint loss Adding them together yields the joint optimization objective loss function. ; S4.5: Utilizing the joint optimization objective loss function Perform backpropagation to calculate the gradient of the U-Net-based convolutional neural network segmentation model and update the model parameters until the model converges.

[0011] Furthermore, the specific steps of S5 are as follows: S5.1: Input the test sample into the trained model to obtain a two-dimensional segmentation probability map under three views; S5.2: Use the view transformation operation to align the two-dimensional segmentation probability maps under the three views to a unified three-dimensional coordinate system to obtain the aligned three-dimensional probability volume under the three views. S5.3: Traverse each voxel in the test sample Calculate voxels Average weighted fusion probability ; S5.4: Binarize the average weighted fusion probability to obtain the three-dimensional prediction result. This allows us to obtain three-dimensional segmentation results for breast tumors.

[0012] The beneficial effects of this invention are: (1) This invention proposes a data augmentation strategy for breast tumors using DCE-MRI (Dynamic Contrast-Enhanced Magnetic Resonance Imaging). By decomposing three-dimensional voxel samples into two-dimensional slices with multiple views, the number of training samples is greatly increased. This strategy effectively alleviates the limitation of the scarcity of medical imaging data on the capabilities of deep learning models, enabling the system to achieve efficient and high-performance tumor segmentation even with limited clinical sample data.

[0013] (2) This invention introduces a dual spatial smoothing constraint mechanism, which is both global and local. Locally, the environmental information between slices is captured by the context-aware module, while globally, it is captured by multi-view... Figure 1 Consistency and spatial similarity constraints enforce topological consistency, thereby ensuring that the reconstructed 3D segmentation boundary has excellent geometric continuity and overall smoothness.

[0014] (3) The cross-view proposed in this invention Figure 1 The consistent segmentation model has been fully validated on multiple authoritative benchmark datasets. Quantitative evaluation results show that the algorithm proposed in this invention significantly outperforms various existing advanced segmentation models in key evaluation metrics such as Dice similarity coefficient (DSC) and average surface distance (ASD), providing more accurate anatomical descriptions of lesions and offering reliable support for clinical diagnosis and decision support systems. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0016] Figure 1 This is a diagram of the overall framework structure of the present invention; Figure 2 These are the smoothing analysis results of the model segmentation results of this invention, where (a) is a visualization of the smoothness of multi-view samples, (b) is a visualization of the smoothness of spatial similarity constraints, and (c) is a visualization of the smoothness of multi-view samples. Figure 1 Visualization of consistency constraint smoothness. Detailed Implementation

[0017] 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.

[0018] like Figure 1 As shown, this embodiment of the invention provides a cross-view... Figure 1 A three-dimensional segmentation method for homogeneous breast tumors includes the following steps: The technical solution adopted in this invention is a cross-view... Figure 1 A three-dimensional segmentation method for homogeneous breast tumors includes the following steps: S1: Acquiring dynamic contrast-enhanced magnetic resonance three-dimensional voxel sample data and the corresponding real segmentation mask Y Slicing operations are performed along three mutually perpendicular orthogonal views to obtain two-dimensional slice sequences and corresponding two-dimensional real segmentation mask sequences under axial, sagittal, and coronal views, respectively. In a specific embodiment of this application, step S1 is as follows: S1.1: Acquiring dynamic contrast-enhanced magnetic resonance three-dimensional voxel sample data and the corresponding real segmentation mask Y; S1.2: Transfer the three-dimensional voxel sample data Slicing operations are performed along three mutually perpendicular orthogonal views, namely the axial, sagittal, and coronal planes, to generate a two-dimensional slice sequence under the three views, as shown in Equation (1), for the actual segmentation mask. Y Perform the same slicing operation to obtain the two-dimensional real mask sequence under the corresponding view, as shown in equation (2): (1) in: A two-dimensional slice in an axial view. For the sagittal plane Figure 2 Fiber slices, A two-dimensional slice in coronal view. For the real number field, For the height of the data, For the width of the data, For the depth of the data, For slicing operations, For the index of the axial section, is the index of the sagittal slice, and z is the index of the coronal slice. This is three-dimensional voxel sample data.

[0019] (2) in: The first axial view Zhang 2D real mask slice The first sagittal view Zhang 2D real mask slice The first coronal view z Zhang 2D real mask slice.

[0020] The obtained two-dimensional real segmentation mask sequences under each view are used as supervision signals. They are used to calculate the segmentation loss function with the segmentation probability map predicted by the U-Net-based convolutional neural network segmentation model to guide the optimization of network parameters.

[0021] Through this step, three-dimensional voxel sample data It is broken down into a huge set of two-dimensional images, which fundamentally alleviates the problem of scarce medical imaging data.

[0022] S2: Perform local context-aware stitching on the two-dimensional slice sequences under the three views respectively to obtain the enhanced slice features under each view; S2.1: For each view, traverse each slice in its two-dimensional slice sequence, define the current slice as the target slice, and extract its adjacent upper and lower slices; S2.2: The target slice is stitched and padded with its adjacent slices above and below it in the channel dimension to obtain the stitched multi-channel data; In a specific embodiment of this application, each slice in a two-dimensional slice sequence under any view is processed, and the slice currently undergoing feature extraction and segmentation prediction is defined as the target slice. For the currently located target slice (let its index in the slice sequence be...),... ), extract its adjacent previous slice by index (index is ) and the next slice (index is The target slice is then stitched together with its adjacent slices above and below it along the channel dimension to form a shape like... Three-channel input data; In addition, to address the issue of information loss in orthogonal dimensional space after two-dimensional slicing, this application proposes a local context-aware stitching strategy. For any target slice, it is stitched together with the two adjacent slices above and below it in the channel dimension. For the first and last slices in the sequence, since they lack adjacent slices, a padding strategy with a value of 1 (1-padding) is used to fill them in.

[0023] S2.3: Construct a residual convolution module. Input the concatenated multi-channel data into the residual convolution module for feature extraction to obtain enhanced slice features containing local spatial context information, i.e., for any view... The extraction process of the residual convolution module is shown in equation (3); (3) in: For view The extracted enhanced slice features containing local spatial context information, For the first Layer convolution operation, This is the sliced ​​data after channel splicing. This is the first layer of convolution operation. This is the second convolution operation. This is the third layer of convolution operation.

[0024] In the specific implementation of this application, the residual convolution module includes three convolutional layers. The first two layers adopt a residual connection structure to alleviate the gradient vanishing problem during deep network training and enhance the effective transmission capability of feature information. The third layer is used to further fuse channel information and output the final context-aware features. Through this structural design, the parameter scale is effectively controlled while ensuring the model's expressive power, thereby improving the stability and robustness of feature extraction. This step combines the target slice with its two adjacent slices above and below it in the channel dimension and fuses them using convolution. This allows the two-dimensional slice, which originally only contained information in one plane, to additionally obtain information from adjacent layers perpendicular to that plane. This step supplements the two-dimensional slice with spatial information in orthogonal directions at the local level.

[0025] S3: Construct a convolutional neural network segmentation model and initialize its parameters. The convolutional neural network segmentation model shares the same weights among the three views. Input the enhanced slice features under the three views into the convolutional neural network segmentation model respectively to obtain the two-dimensional segmentation probability map under the corresponding view. S3.1: Construct a convolutional neural network segmentation model based on the U-Net architecture. The convolutional neural network segmentation model includes an encoder, a decoder, and an output layer. S3.2: Input the enhanced slice features from the axial, sagittal, and coronal views into the convolutional neural network segmentation model to obtain two-dimensional segmentation probability maps from the axial, sagittal, and coronal views, respectively.

[0026] In a specific embodiment of the present invention, in order to improve computational efficiency and significantly reduce the number of model parameters, the enhanced slice features under the three views are processed using the same segmentation network. That is, the enhanced slice features of different views are input sequentially or in parallel into the same network structure, and the convolutional layers and their parameters in the network are completely consistent across different views. This enables the model to have consistent feature extraction and segmentation capabilities under different viewpoints, and outputs two-dimensional segmentation probability maps under axial, sagittal, and coronal views. .

[0027] S4: Calculate the segmentation loss Spatial similarity constraint loss and multiple views Figure 1 Consistency constraint loss The joint optimization objective loss function is used to update the parameters of the convolutional neural network segmentation model, thus obtaining the trained convolutional neural network segmentation model. In a specific embodiment of the present invention, step S4 is as follows: S4.1: Calculate the segmentation loss As shown in equation (4), the segmentation loss Composed of Dice loss and mean squared error loss, as shown in Equations (5) and (6), it takes into account both global region overlap and pixel-level local details, and is used to constrain the consistency between the two-dimensional segmentation probability map output by the model and the two-dimensional real mask under the corresponding view. (4) in: To divide the loss, For Dice's loss, For mean square error loss, This is a two-dimensional segmentation probability map predicted by the model. This refers to a two-dimensional true mask slice at the same location in the same view, corresponding to the two-dimensional segmentation probability map predicted by the model.

[0028] (5) in: For Dice's loss, This represents the total number of pixels in a single slice. For pixel index, For the first The predicted probability of each pixel. For the first A true segmentation mask of 1 pixel, It is a smoothing factor used to avoid the denominator being zero.

[0029] (6) in: This represents the mean square error loss.

[0030] S4.2: Computational spatial similarity constraint loss As shown in Equation (7), a Gaussian-weighted mean square error penalty is applied to the predicted probability maps of adjacent or nearby slices in a slice sequence under the same view to maintain spatial continuity between slices and constrain spatial similarity loss. Based on the assumption of spatial continuity of breast tumors, the segmentation morphology of adjacent slices should be similar, and the similarity decreases as the distance between slices increases. Gaussian weights are used to calculate the mean square error between slices.

[0031] (7) in: This represents the total number of slices for the sample in this view. This is the probability map output by the convolutional neural network segmentation model. For the first Features of each slice For the first Features of each slice To control the Gaussian hyperparameter of the distance weight decay rate, As the normalization factor, For the slice sequence A slice, For the slice sequence m One slice; S4.3: Computing Multi-View Figure 1 Consistency constraint loss As shown in Equation (8), it is used to prevent strong smoothing constraints from causing the model to output a trivial solution where all slices are the same. In the specific embodiments of this application, multiple views Figure 1 Consistency constraint loss The requirement that the reconstructed segmentation results from different orthogonal perspectives maintain topological consistency is expressed as follows: (8) in: For multiple views Figure 1 Consistency constraint loss, For a collection of view pairs, This represents a pair of orthogonal views of axial plane A, sagittal plane S, and coronal plane C. Let be the coordinates of any voxel in three-dimensional space. ) for view In three-dimensional coordinates The predicted probability of the mapping at the location, Represents a three-dimensional spatial domain. The total number of voxels. For in view In three-dimensional coordinates The corresponding predicted probability value, It is the square of the Euclidean distance.

[0032] S4.4: At the global level, to ensure the smoothness of the final 3D reconstruction result and avoid slice discontinuities caused by 2D convolution, this invention will use segmentation loss. Spatial similarity constraint loss and multiple views Figure 1 Consistency constraint loss Adding them together yields the joint optimization objective loss function. ; S4.5: Utilizing the joint optimization objective loss function Perform backpropagation to calculate the gradient of the U-Net-based convolutional neural network segmentation model, and use the optimizer to update the model parameters until the model converges.

[0033] S5: Input the test samples into the trained convolutional neural network segmentation model to obtain the two-dimensional segmentation probability map of each view, align it to the three-dimensional coordinate system, and then perform average weighted fusion and binarization to obtain the three-dimensional breast tumor segmentation result.

[0034] In a specific embodiment of the present invention, step S5 is as follows: S5.1: Input the test samples into the trained model to obtain the two-dimensional segmentation probability map in the axial plane, the two-dimensional segmentation probability map in the sagittal plane, and the two-dimensional segmentation probability map in the coronal plane. S5.2: Use the view transformation operation to align the two-dimensional segmentation probability map under the axial plane, the two-dimensional segmentation probability map under the sagittal plane, and the two-dimensional segmentation probability map under the coronal plane to a unified three-dimensional coordinate system. The calculation formula is shown in Equation (9), and the aligned three-dimensional probability volume under the axial plane, the three-dimensional probability volume under the sagittal plane, and the three-dimensional probability volume under the coronal plane are obtained. (9) in: To align the predicted probability maps of all slices in the axial view to the resulting three-dimensional probability volume in a unified three-dimensional coordinate system after view transformation, This is the 3D probability volume obtained by aligning the predicted probability maps of all slices in the sagittal view after view transformation. This is the three-dimensional probability volume obtained by aligning the predicted probability maps of all slices under the coronal view after view transformation. (.) represents a view transition operation. Two-dimensional segmentation probability map output by a convolutional neural network segmentation model The first axial view Enhanced features of each slice The first sagittal view Enhanced features of each slice The first coronal view Enhanced features of each slice For the index of the axial section, For the index of sagittal slices, For the index of coronal slices, All in the axial view A sequence composed of predicted probability maps of slices. All in the sagittal view A sequence composed of predicted probability maps of slices. All in coronal view A sequence consisting of predicted probability maps of slices.

[0035] The function of S5.2 is to transform the prediction results of the three 2D views into a unified 3D spatial coordinate system, so that the predictions of different views can be aligned at the voxel level, thereby enabling voxel-level averaging fusion and multi-view fusion. Figure 1 Consistency checks provide spatial correspondences.

[0036] S5.3: Traverse each voxel in the test sample Calculate voxels Average weighted fusion probability As shown in equation (10); (10) Among them: voxels The average weighted fusion probability after fusing the three views After conversion to an axial view, in voxels The predicted probability at that location. After conversion to sagittal view, in voxels The predicted probability at that location. After the coronal view is transformed, in the voxel The predicted probability at that location.

[0037] S5.4: The average weighted fusion probability is binarized as shown in Equation (11) to obtain the three-dimensional prediction segmentation mask. This enables three-dimensional segmentation of breast tumors.

[0038] (11) To verify the cross-view-based method proposed in this invention... Figure 1The effectiveness of the consistent 3D breast tumor segmentation method was comprehensively validated experimentally on two publicly available benchmark datasets. In the data preprocessing stage, the input dynamic contrast-enhanced magnetic resonance (DCE-MRI) 3D voxel sample data were single-channel images. This data was obtained by subtracting the pre-contrast image from the first contrasted image, resulting in a difference image. All images and their corresponding ground truth segmentation masks were resampled. Isotropic resolution was used to ensure spatial consistency. Regarding dataset partitioning, the BreastCancer DCE-MRI dataset contained 100 cases, partitioned into 80 cases for training and 20 cases for testing; the MAMA-MIA dataset selected 291 bilateral breast MRI cases, partitioned into 233 cases for training and 58 cases for testing. The experiments used Dice similarity coefficient (DSC), positive predictive value (PPV), sensitivity (SEN), and average surface distance (ASD) as evaluation metrics.

[0039] To evaluate the overall segmentation performance of the model in this invention under normal data volume, it was compared with 10 current mainstream advanced medical image segmentation methods (including methods based on 3D CNN, Transformer, etc.). The experimental results are shown in Table 1.

[0040] Table 1. Performance comparison of different methods on two breast tumor segmentation datasets.

[0041] As shown in Table 1, the method of this invention achieved optimal results in both DSC and ASD metrics on the two datasets. Particularly on the Breast Cancer DCE-MRI dataset, the ASD of this invention was reduced to a minimum of 5.47 mm, indicating that the 3D surface reconstructed by this method exhibits the highest smoothness and geometric continuity. Furthermore, this model achieves a good balance between PPV and SEN, effectively overcoming the serious undersegmentation or oversegmentation problems present in other models.

[0042] To verify the advantages of the multi-view slicing strategy of this invention in expanding the sample space and alleviating data scarcity, experiments were conducted to evaluate the segmentation performance of each model under extremely limited sample conditions (1%, 5%, 10%, 20%, 40%, and 60% of the training data). Some key data comparison results are shown in Table 2 (taking DSC and ASD as examples).

[0043] Table 2. Performance comparison of different methods under different training data ratios (Breast Cancer DCE-MRI dataset)

[0044] As shown in Table 2, under the extremely scarce condition of using only 1% of the training data, the DSC performance of the model of this invention is significantly superior (reaching 61.72%), while most of the comparison methods (such as nnUNet and UXNET) suffer severe performance degradation or overfitting (e.g., TransBTS has an ASD as high as 257.40 mm and its segmentation accuracy collapses to 0). With the increase in data volume, this invention consistently maintains superior performance, demonstrating that the strategy of decomposing three-dimensional voxels into orthogonal two-dimensional slices can expand the effective training space by hundreds of times, fundamentally solving the problem of overfitting in small-sample training.

[0045] To evaluate the clinical deployment potential and computational efficiency of the present invention, a comprehensive comparison was made of the number of parameters, computational complexity (FLOPs), inference time per sample, and DSC of each method. The results are shown in Table 3.

[0046] Table 3 Comparison of model efficiency and performance of different methods

[0047] As shown in Table 3, because this invention employs a parameter-sharing two-dimensional convolutional architecture across the backbone networks of the three views, its model parameter count is only 8.114M, far lower than mainstream Transformer models (such as UNETR's 149.347M). Although some FLOPs are added to fuse multi-view information, through parallel optimization, the single-sample inference time is extremely short (only a few seconds). While ensuring the highest segmentation accuracy (DSC 82.07%), it achieves an extremely lightweight structure and has strong practical deployment capabilities.

[0048] In order to systematically evaluate the core modules (multi-view strategy, consistency constraint loss) in the framework of this invention Spatial similarity constraint loss Hybrid segmentation loss and Based on the individual contribution of ), the present invention designed an ablation experiment, and the results are shown in Table 4.

[0049] Table 4 Ablation experimental results on the Breast Cancer DCE-MRI dataset

[0050] Table 4 shows that removing any of the views (axial, coronal, or sagittal) leads to a decrease in DSC and an increase in ASD, demonstrating that each of the three orthogonal views provides a unique spatial context. Simultaneously, removing the smoothing constraint (

[0051] and The removal of both constraints resulted in a 3.13% decrease in DSC, demonstrating the central role of dual constraints in producing smooth segmentation consistent with anatomical logic. Furthermore, the absence of either constraint alone... or Both can lead to performance degradation, validating the complementary advantage of using them together to simultaneously address global class imbalance and pixel-level boundary refinement.

[0052] To evaluate the actual effect of global spatial smoothing constraints, this invention uses the volume change rate to quantitatively evaluate the segmentation smoothness. The visualization results are as follows: Figure 2 As shown. Combined with Figure 2 As shown in the histograms (a), (b), and (c), compared to ablation variants lacking smoothness constraints or using only single / dual views, the complete model of this invention exhibits a significantly reduced volume change rate, demonstrating smoothness closest to the real segmentation mask. The single-view configuration exhibits the highest NVCR, indicating severe surface unevenness and tortuosity. When spatial similarity constraints are introduced... ( Figure 2 b) and multi-view Figure 1 Consistency constraints ( Figure 2 c) After that, the model is forced to maintain morphological consistency across different viewpoints and adjacent slices, completely eliminating the stepped artifacts and spatial discontinuities caused by 2D slicing. This directly proves that the present invention significantly optimizes 3D geometric smoothness while maintaining high accuracy.

[0053] In summary, this invention not only fundamentally expands the sample size to address the limited sample problem by slicing along three orthogonal views, but also utilizes a local context-aware module to compensate for the spatial information lost in two-dimensional slicing. More importantly, to completely eliminate staircase artifacts and generate smooth three-dimensional structures, this invention introduces global spatial similarity constraint loss and multi-view loss into the constructed loss function. Figure 1 The consistency constraint loss is used to achieve spatial continuity and structural consistency of the 3D segmentation results through joint optimization of the two methods. Compared with traditional methods, the model of this invention can learn more stable and smooth 3D spatial reconstruction characteristics under few sample conditions. Its experimental results are verified in Tables 1 and 2 of the embodiments, which can meet the requirements of clinical applications for segmentation accuracy and structural continuity.

[0054] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A three-dimensional breast tumor segmentation method based on cross-view consistency, characterized in that, Includes the following steps: S1: Acquiring dynamic contrast-enhanced magnetic resonance three-dimensional voxel sample data and the corresponding real segmentation mask Y Slicing operations are performed along three mutually perpendicular orthogonal views to obtain two-dimensional slice sequences and corresponding two-dimensional real segmentation mask sequences under axial, sagittal, and coronal views, respectively. S2: Perform local context-aware stitching on the two-dimensional slice sequences under the three views respectively to obtain the enhanced slice features under each view; S3: Construct a convolutional neural network segmentation model and initialize its parameters. The convolutional neural network segmentation model shares the same weights among the three views. The enhanced slice features under each view are input into the convolutional neural network segmentation model to obtain the two-dimensional segmentation probability map under the corresponding view. S4: Calculate the segmentation loss Spatial similarity constraint loss and multi-view consistency constraint loss The objective loss function is jointly optimized, and the parameters of the convolutional neural network segmentation model are updated using the jointly optimized objective loss function to obtain the trained convolutional neural network segmentation model. S5: Input the test sample into the trained convolutional neural network segmentation model to obtain the three-dimensional breast tumor segmentation result.

2. The three-dimensional breast tumor segmentation method based on cross-view consistency according to claim 1, characterized in that, The specific steps for S1 are as follows: S1.1: Acquiring dynamic contrast-enhanced magnetic resonance three-dimensional voxel sample data and the corresponding real segmentation mask Y ; S1.2: Transfer the three-dimensional voxel sample data Slice along three mutually perpendicular orthogonal views to generate a 2D slice sequence in the three views, and then apply the result to the actual segmentation mask. Y Perform a slicing operation to generate a sequence of two-dimensional real segmentation masks in three views.

3. The three-dimensional breast tumor segmentation method based on cross-view consistency according to claim 1, characterized in that, The specific steps for S2 are as follows: S2.1: For each view, traverse each slice in its two-dimensional slice sequence, define the current slice as the target slice, and extract its adjacent upper and lower slices; S2.2: The target slice is stitched and padded with its adjacent slices above and below it in the channel dimension to obtain the stitched multi-channel data; S2.3: Construct a residual convolution module, input the spliced ​​multi-channel data into the residual convolution module for feature extraction, and obtain the enhanced slice features under the three views respectively.

4. The three-dimensional breast tumor segmentation method based on cross-view consistency according to claim 1, characterized in that, The specific steps for S3 are as follows: S3.1: Construct a convolutional neural network segmentation model based on the U-Net architecture; S3.2: Input the enhanced slice features from the three views into the convolutional neural network segmentation model to obtain two-dimensional segmentation probability maps from the three views respectively.

5. A three-dimensional breast tumor segmentation method based on cross-view consistency according to claim 1, characterized in that, The specific steps for S4 are as follows: S4.1: Define the segmentation loss As shown in Equation (1), the consistency between the two-dimensional segmentation probability map output by the model and the two-dimensional true segmentation mask under the corresponding view is constrained, and the segmentation loss is... It is composed of Dice loss and mean square error loss, as shown in Equation (2) and Equation (3) respectively; （1） in: To divide the loss, For Dice's loss, For mean square error loss, This is a two-dimensional segmentation probability map predicted by the model. A two-dimensional true segmentation mask slice at the same location in the same view, corresponding to the two-dimensional segmentation probability map predicted by the model; （2） in: This represents the total number of pixels in a single slice. For pixel index, For the first The predicted probability of each pixel. For the first A true segmentation mask of 1 pixel, It is a smoothing factor; （3） S4.2: Define spatial similarity constraint loss As shown in equation (4); （4） in: This represents the total number of slices for the sample in this view. This is the probability map output by the convolutional neural network segmentation model. For the first Features of each slice For the first Features of each slice To control the Gaussian hyperparameter of the distance weight decay rate, As the normalization factor, For the slice sequence A slice, For the slice sequence m One slice; S4.3: Define the loss due to multi-view consistency constraints As shown in equation (5), it is used to prevent strong smoothing constraints from causing the model to output a trivial solution where all slices are the same. （5） in: For multi-view consistency constraint loss, For a collection of view pairs, Let be the coordinates of any voxel in three-dimensional space. ) for view In three-dimensional coordinates The predicted probability of the mapping at the location, Represents a three-dimensional spatial domain. The total number of voxels. For in view In three-dimensional coordinates The corresponding predicted probability value, The square of the Euclidean distance; S4.4: The segmentation loss Spatial similarity constraint loss and multi-view consistency constraint loss Adding them together yields the joint optimization objective loss function. ; S4.5: Utilizing the joint optimization objective loss function Perform backpropagation to calculate the gradient of the U-Net-based convolutional neural network segmentation model and update the model parameters until the model converges.

6. The three-dimensional breast tumor segmentation method based on cross-view consistency according to claim 1, characterized in that, The specific steps for S5 are as follows: S5.1: Input the test sample into the trained model to obtain a two-dimensional segmentation probability map under three views; S5.2: Use the view transformation operation to align the two-dimensional segmentation probability maps under the three views to a unified three-dimensional coordinate system to obtain the aligned three-dimensional probability volume under the three views. S5.3: Traverse each voxel in the test sample Calculate voxels Average weighted fusion probability ; S5.4: Binarize the average weighted fusion probability to obtain the 3D segmentation result. This allows us to obtain three-dimensional segmentation results for breast tumors.

Images