A synthetic molybdenum target pseudo-label guided three-dimensional breast MRI segmentation method

Abstract

This invention discloses a three-dimensional breast MRI segmentation method guided by synthetic mammogram pseudo-labels. This method aims to solve the problem of automatic 3D breast segmentation caused by the scarcity of voxel-level annotations in 3D MRI and the significant non-rigid deformation between real 2D mammograms and 3D MRI images, making direct utilization difficult. The method first converts the 3D MRI image into a 2D synthetic mammogram image through physical simulation projection. Then, using a small number of labeled synthetic images, a 2D teacher network is trained through feature decoupling and uncertainty filtering mechanisms, while high-quality 2D pseudo-labels are generated for a large number of unlabeled synthetic images. Finally, a 3D student network is constructed, projecting the 3D segmentation prediction onto a 2D plane through a differentiable projection layer. Using the 2D pseudo-labels as weak supervision signals and combined with spatial smoothing constraints, the network is trained to achieve high-precision automatic 3D breast segmentation. This invention significantly reduces annotation dependence, avoids cross-modal registration difficulties, and produces 3D segmentation results with clear boundaries and spatial continuity.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and computer-aided diagnosis technology, specifically relating to a three-dimensional breast MRI segmentation method guided by a synthetic mammogram pseudo-label. Background Technology

[0002] Breast cancer is the most common malignant tumor among women worldwide, making its early detection and accurate diagnosis crucial. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), with its advantages such as high soft tissue resolution and no ionizing radiation, can provide three-dimensional anatomical and blood flow information, and has become an important imaging tool for breast cancer screening and assessment. Achieving precise segmentation of breast tissue is a key prerequisite for computer-aided diagnosis, calculating glandular density, and locating lesions.

[0003] However, existing technical solutions face the following main challenges:

[0004] 1) Fully supervised methods have extremely high annotation costs: High-precision 3D segmentation models based on fully convolutional neural networks rely on a large amount of pixel-level (voxel-level) labeled data. Due to the large number of 3D MRI data slices and the varied morphology and blurred boundaries of breast tissue, it is time-consuming and laborious for radiologists to manually delineate each layer in detail, resulting in high annotation costs and subjective inconsistencies. This seriously hinders the construction of large-scale, high-quality datasets and the clinical promotion of fully supervised models.

[0005] 2) Insufficient segmentation accuracy of weakly supervised methods: In order to reduce the annotation requirements, weakly supervised learning methods attempt to use low-cost annotation methods such as image-level labels and bounding boxes. However, since the breast is a non-rigid soft tissue, the gray-level contrast with surrounding tissues (such as fat and chest wall muscles) is not obvious in some sequences. These weakly supervised signals are difficult to capture fine edge information, resulting in problems such as blurred boundaries, undersegmentation, or overflow in the segmentation results, which cannot meet the accuracy requirements of clinical quantitative analysis.

[0006] 3) The inability to directly utilize cross-modal data (2D-3D gap): Clinically, there is a vast amount of low-cost, easily accessible 2D mammography images and annotations. Theoretically, using this 2D information to guide 3D MRI segmentation has enormous potential, but it faces fundamental technical obstacles.

[0007] The imaging mechanisms are different: MRI is 3D tomographic imaging, while mammography is 2D perspective projection, and the two belong to different modalities.

[0008] There is significant non-rigid deformation: during mammography, the breast is compressed and flattened by a splint; while during MRI scans, the breast naturally droops. This large, non-linear geometric deformation between the two postures makes accurate spatial registration between real 2D mammograms and 3D MRI data virtually impossible.

[0009] Therefore, how to overcome the huge deformation differences across modalities and effectively utilize readily available 2D annotation information to drive high-precision automatic 3D MRI breast segmentation in the absence of large-scale 3D voxel-level annotation is a key technical challenge that urgently needs to be solved in the field of medical image analysis. This invention aims to provide an innovative solution to this problem. Summary of the Invention

[0010] To address the aforementioned problems, the present invention aims to provide a method for three-dimensional breast MRI segmentation guided by synthetic mammogram pseudo-tags.

[0011] The specific technical solution is as follows:

[0012] A method for three-dimensional breast MRI segmentation guided by synthetic mammography pseudotags includes the following steps:

[0013] S1. Data Preprocessing and Standardization: Receive raw 3D MRI volumetric data and perform image correction and resampling;

[0014] S2. Cross-modal image generation and enhancement: Establish the mapping relationship from 3D space to 2D plane, convert the preprocessed 3D data into 2D synthetic images through physical simulation projection algorithm, and perform contrast enhancement and view normalization;

[0015] S3. Semi-supervised 2D pseudo-label construction based on feature decoupling: The synthesized 2D image is divided into a labeled set and an unlabeled set. The supervision signal is decomposed into a subject mask and an edge mask using the feature decoupling module to train a multi-task 2D teacher network. An uncertainty filtering mechanism is introduced to generate high-confidence 2D pseudo-labels for the unlabeled set.

[0016] S4. 3D Weakly Supervised Training Based on Projection Consistency and Spatial Continuity: A 3D student network is established, a parameter-free differentiable projection layer is introduced at the output, the projection consistency loss is calculated using 2D pseudo-labels, and the spatial smoothness loss is combined to guide the training of the 3D network.

[0017] S5. Model Inference and Post-processing: Full-graph inference is performed using the trained 3D student network, and the final 3D segmentation result is obtained through maximum connected component analysis and edge smoothing.

[0018] Furthermore, S1 specifically includes:

[0019] S101: N4 bias field correction. Original MRI images are often affected by magnetic field inhomogeneity, exhibiting low-frequency intensity variations, resulting in significant differences in grayscale values ​​of the same tissue at different locations in the image. This invention uses the N4 algorithm to estimate and remove this bias field.

[0020] Assuming the observed MRI images Images of real anatomical structures It consists of a smooth bias field B and additive noise N, and its mathematical imaging model can be expressed as:

[0021]

[0022] in, Represents the spatial coordinates of a voxel.

[0023] This step estimates the bias field using an iterative optimization algorithm. The distribution of the image is determined and removed from the observed image to recover the corrected image. :

[0024]

[0025] This step significantly improves the grayscale consistency of soft tissues in the image.

[0026] S102: Image resampling. Raw MRI data often has anisotropic spatial resolution, which is not conducive to the extraction of spatial features by 3D convolutional neural networks.

[0027] This step employs a high-order interpolation algorithm to resample the S101-corrected image to isotropic resolution. The resampled target image... At the new coordinates The formula for calculating pixel values ​​is as follows:

[0028]

[0029] in:

[0030] These are discrete neighborhood voxels in the original image coordinate system;

[0031] This is the interpolation kernel function used to calculate the weighted contribution of neighborhood voxels to the target point.

[0032] Furthermore, S2 specifically includes:

[0033] S201: Physical Simulation Projection

[0034] To simulate the imaging process of X-rays penetrating breast tissue, this invention establishes a projection mapping in the coronal plane. Let the preprocessed 3D MRI volumetric data be... The z-axis corresponds to the direction of the human body from head to toe, and the total number of slices is N. z .

[0035] This step uses the mean density projection algorithm to simulate the cumulative attenuation effect of X-rays, preserving the overall structure of soft tissue. The corresponding formula is as follows:

[0036]

[0037] S202: Limiting Contrast-Adaptive Histogram Equalization

[0038] The original projected image typically has a concentrated grayscale distribution and low contrast. This step uses the CLAHE algorithm to enhance the distinction between breast glandular tissue and adipose tissue.

[0039] The algorithm divides the image into non-overlapping patches and calculates local histograms. A cropping threshold is introduced to prevent noise amplification. The local histogram is trimmed and reshaped, resulting in the trimmed histogram. satisfy:

[0040]

[0041] Excess pixels are evenly distributed across the rest of the histogram, and finally, bilinear interpolation is used to eliminate blockages, resulting in an enhanced image. .

[0042] S203: View Standardization

[0043] To eliminate the interference of different patient positions and differences between the left and right breasts on network training, the enhanced images output by S202 need to be processed. Perform geometric standardization.

[0044] Segmentation: Based on the geometric center Cut into left breast view and right breast view .

[0045] Flip: The standard anatomical orientation is defined as "nipple to the right, chest wall to the left". If the image orientation does not conform to the standard, a horizontal flip is performed. Let the image width be W, and the flip operation is defined as:

[0046]

[0047] This step yields the final standardized dataset used for training.

[0048] Furthermore, S3 specifically includes:

[0049] S301: Label Feature Decoupling Based on Distance Transformation

[0050] This invention abandons traditional binary mask supervision and instead utilizes Euclidean distance transformation to convert binary labels M into soft labels containing rich spatial information, constructing a subject probability map. To emphasize the feature stability of the breast center region, a centrality-based soft label is constructed. First, the binary mask M is Gaussian smoothed, and then the Euclidean distance from the foreground pixel to the background boundary is calculated. To expand the high-confidence region, the distance map is nonlinearly mapped and normalized.

[0051]

[0052] generated The distribution is mountain-shaped, with the central probability approaching 1 and the periphery approaching 0, effectively guiding the network to focus on the core area of ​​the subject.

[0053] Next, an edge probability map is constructed. This is a composite distance metric defined to address the potential boundary truncation problem of breast tissue in mammograms. First, the distance from each pixel to the background contour is calculated. And the minimum distance from the pixel to the four physical boundaries of the image. The effective boundary distance is defined as the minimum of the two. :

[0054]

[0055] Then, an inverse distance weighting function is used to generate an edge map, so that pixels closer to the boundary have higher response values:

[0056]

[0057] in For distance control threshold, To prevent minute division by zero, this edge map not only highlights the tissue outline but also explicitly encodes the truncated regions of the image edges.

[0058] S302: Construction of Composite Loss Function

[0059] This invention constructs a 2D teacher network that generates a single-channel prediction probability map P at the output. A multi-dimensional composite loss function is designed in this invention. Using binary truth masks and edge weight map The predicted graph P is jointly supervised, and the composite loss function is defined as follows:

[0060]

[0061] The specific functions of each component are as follows:

[0062] Dice coefficient loss: Calculating the prediction graph With binary truth mask The set similarity between them. This is the main segmentation loss, used to constrain the overall shape of the prediction results to be consistent with the gold standard.

[0063] Binary cross-entropy loss: Calculation of prediction graph With the main probability diagram The differences between them. Due to Including center confidence information based on distance transformation, this loss forces the network to learn the internal probability distribution, enhancing internal consistency.

[0064] Normalized Boundary Cross-Union Loss: The calculation of this loss function depends on the prediction graph. Binary truth mask and edge weight map .in, Used to determine the reference region for segmentation, As a spatial weighting term, higher weights are assigned to intersection and union calculations near the boundary. This forces the network to focus on optimization. The defined contour region, and by Further enhance the sharpness of the edges.

[0065] S303: Pseudo-label reasoning and uncertainty filtering

[0066] Using a converged 2D teacher network trained with S302, inference is performed on synthetic images in an unlabeled dataset, and a predicted probability map P is output.

[0067] To prevent low-quality prediction noise from interfering with subsequent 3D network training, this invention introduces a dual-threshold uncertainty filtering mechanism, dividing the prediction results into a definite foreground, a definite background, and an uncertain, ignored region. The generated final pseudo-labels... The definition is as follows:

[0068]

[0069] Furthermore, S4 specifically includes:

[0070] S401: Forward propagation inputs 3D MRI data into a 3D Student Network. After 3D convolution operations, the network generates a 3D probability map at the output layer through Softmax normalization, representing the probability that each voxel belongs to breast tissue.

[0071] S402: Microprojection

[0072] By introducing a parameter-free differentiable projection layer, the mean value is projected along the coronal plane to generate a 2D projection prediction map. :

[0073]

[0074] This step reduces the dimensionality of the 3D segmentation so that it can be compared with the 2D pseudo-labels.

[0075] S403: Calculation of Mixed Losses

[0076] Total loss It consists of two parts: projection consistency loss and spatial smoothing loss. The projector execution loss is used to calculate the Dice and BCE losses of the 2D projection prediction map and the mixed supervision signal to constrain the projection shape of the 3D structure to be consistent with the molybdenum target. The spatial smoothing loss directly applies a total variation constraint in the Z-axis direction to the 3D probability map to penalize drastic abrupt changes between slices and prevent the generation of "cylinder" artifacts.

[0077] S404: Backpropagation

[0078] Calculate total loss The gradient is backpropagated through the projection layer to the 3D student network to update the model parameters.

[0079] Furthermore, S5 specifically includes:

[0080] S501: Full-Graph Reasoning and Binarization

[0081] 3D MRI data is input into the network, which outputs a 3D probability prediction map. A threshold is set, and the probability map is converted into an initial binary segmentation mask to distinguish the foreground breast tissue from the background.

[0082] S502: Extraction of 3D Maximum Connected Components

[0083] 3D connected component analysis is performed on the binary mask. Since the breast is usually the largest continuous tissue in the image, this step calculates the number of voxels in all connected regions, retains only the largest connected region as the main breast tissue, and automatically filters out free false positive noise in the background.

[0084] S503: 3D Morphological Smoothing

[0085] To obtain smoother boundaries that better conform to anatomical structures, 3D morphological closing operations are performed on the preserved connected regions, first dilating and then eroding. This operation effectively fills tiny voids within the segmented region and smooths jagged edges on the object's surface, outputting a final high-precision 3D segmentation result.

[0086] The beneficial effects of this invention are as follows:

[0087] 1) It overcomes the challenge of cross-modal registration and significantly reduces annotation costs. By establishing a differentiable mapping between 3D MRI and 2D synthetic mammograms through physical simulation projection and differentiable projection layers, it cleverly avoids the registration problem caused by non-rigid deformation between real mammograms and MRI. At the same time, it uses easily obtainable 2D pseudo-labels to replace expensive 3D voxel-level annotations for weakly supervised training, achieving 3D segmentation performance comparable to fully supervised methods while significantly reducing the workload of doctors (reducing time costs by more than 90%).

[0088] 2) Significantly improves the sharpness and accuracy of segmentation results at tissue boundaries, effectively solving the problems of edge blurring and region truncation. In the process of constructing 2D pseudo-labels, this invention introduces a feature decoupling module and a neighborhood-optimized intersection-over-union (NBIOU) loss function. By decoupling the supervision signal into subject probabilities and edge weights, and combining this with a dual-threshold uncertainty filtering mechanism, this method can effectively identify and remove noise and artifacts. This strategy allows the model to accurately capture high-frequency contour features while focusing on internal structural consistency. Through these techniques, this invention solves the common problems of "undersegmentation" or "boundary overflow" in traditional methods, significantly improving the sharpness and accuracy of segmentation results at tissue boundaries, and providing a more reliable and refined solution for 3D breast MRI segmentation.

[0089] 3) It ensures the spatial continuity and anatomical rationality of the 3D segmentation results. Addressing the "cylinder artifact" problem (i.e., identical Z-axis slice predictions) that easily arises in weakly supervised projection methods, this invention introduces a total variational spatial smoothing loss during 3D training, directly applying inter-layer constraints to the 3D probability map and combining it with maximum connected component post-processing. This not only ensures that the generated 3D breast volume is spatially smooth and continuous but also automatically filters out free noise in the background, outputting a high-quality segmentation mask that conforms to clinical anatomy. Attached Figure Description

[0090] Figure 1 This is a flowchart illustrating the overall process of a synthetic mammogram pseudo-label-guided three-dimensional breast MRI segmentation method according to the present invention.

[0091] Figure 2 This is a detailed flowchart of the 2D pseudo-label construction module based on feature decoupling in this invention;

[0092] Figure 3 This is a schematic diagram of the architecture of the 3D training module based on projection consistency and spatial continuity in this invention. Detailed Implementation

[0093] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited thereto.

[0094] like Figure 1 As shown, a method for three-dimensional breast MRI segmentation guided by synthetic mammography pseudo-tags is described below:

[0095] S1. Data Preprocessing and Standardization

[0096] This step aims to establish a unified data benchmark to provide high-quality input for subsequent processing.

[0097] S101: N4 bias field correction: The N4 algorithm is used to automatically estimate and remove low-frequency magnetic field inhomogeneity artifacts in MRI images, correcting spatial deviations in tissue intensity.

[0098] S102: Image Resampling: To address the anisotropy (large interslice spacing) issue in the original MRI data, a B-spline interpolation algorithm is used to resample voxels to a standardized isotropic resolution (1.0 × 1.0 × 1.0 mm in this embodiment). 3 ), to match the input requirements of 3D networks.

[0099] S103: ROI extraction: Otsu thresholding was used to remove background air, and morphological operations were combined to remove the pectoralis major muscle and thoracic organs, retaining only the region of interest containing breast tissue.

[0100] S2. Cross-modal image generation and enhancement

[0101] This step establishes a mapping from 3D MRI to 2D mammography through physical simulation.

[0102] S201: Physical Simulation Projection: Simulating the principle of X-ray penetration imaging, 3D data is projected along the coronal plane. This embodiment preferably uses Average Intensity Projection (AIP), which simulates the cumulative attenuation effect of X-rays by calculating the average value of voxels along the light path, thereby generating a 2D synthetic image that preserves the overall structure of soft tissue. It should be noted that this invention can also be combined with Maximum Intensity Projection (MIP) to capture features of dense glands.

[0103] S202: Enhancement and Normalization: Limit Contrast Adaptive Histogram Equalization (CLAHE) is applied to the projected image to enhance contrast; then the left and right breasts are segmented according to the geometric center, and the right breast view is uniformly flipped to the standard anatomical orientation.

[0104] S3. Semi-supervised 2D pseudo-label construction based on feature decoupling

[0105] like Figure 2 As shown, this step uses a small amount of labeled data to train the teacher network and generate high-quality pseudo-labels for the unlabeled data.

[0106] S301: Label Feature Decoupling: Using Euclidean distance transform, binary labels are converted into soft labels containing spatial information.

[0107] Main probability diagram ( ): Generated based on the distance from pixel to background, with values ​​decreasing from the center of the breast to the edge, representing the center confidence;

[0108] Edge weight map ( ): Generated based on the reciprocal of the minimum distance from the pixel to the background and image boundary. The value is extremely high at the contour and represents the edge strength.

[0109] S302: Construction of Composite Loss Function: Construct a 2D teacher network output single-channel prediction graph P, and perform multi-objective composite loss function construction. Supervision:

[0110]

[0111] in, Constrain the internal probability distribution, Use edge weighting maps to specifically penalize boundary errors.

[0112] S303: Pseudo-label Inference and Filtering: Utilizing a teacher network to infer from unlabeled data, a dual-threshold uncertainty filtering mechanism is introduced to generate the final pseudo-labels. :

[0113]

[0114] The uncertain regions in the middle are automatically ignored in subsequent training.

[0115] S4. 3D Weakly Supervised Training Based on Projection Consistency and Spatial Continuity

[0116] like Figure 3 As shown, this step enables cross-dimensional learning from 2D labels to 3D segmentation.

[0117] S401: 3D Forward Propagation: Construct a 3D student network (this embodiment uses the V-Net architecture), input 3D MRI data, and output a 3D probability map representing voxel-level classification probabilities. .

[0118] S402: Differentiable Projection: Introducing a parameter-free differentiable projection layer to... A 2D projection prediction map is generated by performing mean projection along the Z-axis. Establish gradient transmission pathways between 2D and 3D spaces.

[0119] S403: Mixed Loss Calculation: Defining the Total Loss Function To simultaneously constrain the projected shape and spatial continuity:

[0120]

[0121] in, Includes Dice and BCE losses for fitting 2D pseudo-labels; Using total variational (TV) regularization, penalizing Drastic abrupt changes between Z-axis slices prevent the formation of interlayer discontinuity artifacts.

[0122] S5. Model Inference and Post-processing

[0123] S501: Full-map reasoning: Using a trained 3D student network, full-map reasoning is performed on the test set MRI data, and a 3D probability map is output. In this embodiment, the binarization threshold is set to 0.5, that is, when the voxel prediction probability is greater than or equal to 0.5, it is marked as foreground (breast tissue), and when it is less than 0.5, it is marked as background.

[0124] S502: Post-processing: Perform 3D maximum connected component extraction, retain only the largest connected component as the main body of the breast, and automatically filter out free noise in the background; finally, use 3D morphological closing operation to smooth the segmentation boundary and output the final result.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for three-dimensional breast MRI segmentation guided by synthetic mammography pseudo-tags, characterized in that, Includes the following steps: S1. Preprocess and standardize the raw 3D MRI volumetric data; S2. By physically simulating projection, the preprocessed and standardized 3D MRI volume data are converted into synthetic 2D mammograms, and image enhancement and view normalization are performed. S3. Semi-supervised pseudo-label construction based on feature decoupling: The synthesized 2D molybdenum target image is divided into a labeled set and an unlabeled set. Using the feature decoupling module, the binary label mask of the labeled set is decoupled into a main probability map and an edge weight map. Based on this, a 2D teacher network is trained. The trained 2D teacher network is used to infer the images in the unlabeled set, and combined with an uncertainty filtering mechanism, 2D pseudo-labels are generated for the unlabeled set. S4. 3D Weakly Supervised Training Based on Projection Consistency and Spatial Continuity: A 3D student network is constructed, whose input receives preprocessed and standardized 3D MRI volume data, and outputs a 3D probability map. The 3D probability map is projected along the coronal plane into a 2D projection prediction map through a parameter-free differentiable projection layer. Using 2D pseudo-labels as supervision signals, the projection consistency loss between the 2D projection prediction map and the 2D pseudo-labels is calculated. A spatial smoothing loss is applied to the 3D probability map. The 3D student network is trained in a weakly supervised manner by combining the projection consistency loss and the spatial smoothing loss. S5. Model Inference and Post-processing: Input the 3D MRI data to be segmented into the trained 3D student network to obtain the initial 3D segmentation results, and then perform post-processing to obtain the final 3D breast segmentation mask.

2. The method for three-dimensional breast MRI segmentation guided by synthetic mammography pseudo-tags as described in claim 1, characterized in that, Step S1 includes: S101. The N4 bias field correction algorithm is used to correct the intensity inhomogeneity of the original 3D MRI volume data; S102. Resample the corrected image to an isotropic resolution.

3. The method for three-dimensional breast MRI segmentation guided by synthetic mammography pseudo-tags as described in claim 1, characterized in that, Step S2 includes: S201. Project the preprocessed and standardized 3D MRI volume data along the coronal plane using mean density to generate an initial 2D composite image; S202. Perform contrast-limited adaptive histogram equalization on the initial 2D composite image to enhance contrast; S203. Perform view normalization on the contrast-enhanced image, including: dividing the image into a left breast view and a right breast view based on the geometric center of the image, and unifying the view orientation to the standard anatomical orientation.

4. The method for three-dimensional breast MRI segmentation guided by synthetic mammography pseudo-tags as described in claim 1, characterized in that, Step S3 includes: S301. Feature decoupling: Perform distance transformation on the binary labeled mask with a labeled set to generate a main probability map representing the confidence of the central region of the breast, and an edge weight map representing the tissue contour and the image boundary region. S302. 2D Teacher Network Training: The loss function of the 2D teacher network is a composite loss function, including: Dice loss based on binary labeled mask, binary cross-entropy loss based on the main probability graph, and normalized boundary intersection-union ratio loss based on binary labeled mask and edge weight graph; S303. Pseudo-label generation and filtering: The trained 2D teacher network is used to infer the unlabeled set of images to obtain a predicted probability map; a high threshold and a low threshold are set, and pixels with predicted probabilities higher than the high threshold are judged as foreground, pixels with predicted probabilities lower than the low threshold are judged as background, and pixels in between are marked as ignored areas, thereby generating filtered 2D pseudo-labels.

5. The method for three-dimensional breast MRI segmentation guided by synthetic mammography pseudo-tags as described in claim 1, characterized in that, Step S4 includes: S401. Input 3D MRI data into a 3D student network and obtain a 3D probability map through forward propagation; S402. The 3D probability map is averaged and projected along the coronal plane using a parameter-free differentiable projection layer to obtain a 2D projection prediction map. S403. Calculate the hybrid loss: The projection consistency loss is calculated based on the 2D projection prediction map and the 2D pseudo-label; the spatial smoothing loss is the total variation regularization loss applied to the 3D probability map in the slice direction to penalize inter-layer abrupt changes; the total loss is the weighted sum of the projection consistency loss and the spatial smoothing loss. S404. Update the parameters of the 3D student network through backpropagation based on the total loss.

6. The method for three-dimensional breast MRI segmentation guided by synthetic mammography pseudo-tags as described in claim 1, characterized in that, Step S5 includes: S501. Threshold binarize the 3D probability map output by the 3D student network to obtain the initial binary segmentation mask; S502. Perform three-dimensional connected component analysis on the initial binary segmentation mask, extract and retain the connected component with the largest volume as the main body of the breast; S503. Perform a three-dimensional morphological closing operation on the main body of the breast to smooth the segmentation boundaries and fill the internal cavities, thus obtaining the final segmentation result.

Images