A semi-supervised multi-organ image segmentation method for solving class imbalance and distribution bias

Abstract

This invention provides a semi-supervised multi-organ image segmentation method to address class imbalance and distribution shift, comprising the following steps: S1, acquiring an abdominal multi-organ medical image dataset and converting all images into uniform-sized three-dimensional voxel tensors; S2, constructing a dual-headed decoupled three-dimensional neural network model, outputting the trained dual-headed network model and corresponding class temperature parameters; S3, applying the trained dual-headed network model to the input tensor to generate multi-organ segmentation prediction results, and simultaneously calculating the segmentation confidence scores for each class; S4, identifying the difficult classes with low confidence scores based on the segmentation confidence scores of each class; generating synthetic samples to enhance the minority class; S5, outputting the trained dual-headed decoupled three-dimensional neural network model; S6, outputting the corresponding multi-organ segmentation results. This invention significantly improves the overall accuracy of multi-organ segmentation and the performance of small organ recognition by dynamically calibrating class distribution shift and guiding the generation of enhancements for the minority class.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more particularly to a semi-supervised multi-organ image segmentation method for solving class imbalance and distribution shift. Background Technology

[0002] Abdominal multi-organ medical image segmentation is a key task in medical image analysis. Its goal is to automatically process input abdominal CT or MRI images and accurately identify and segment regions of multiple organs, such as the liver, pancreas, kidneys, spleen, and stomach. In clinical practice, these segmentation results are widely used in disease diagnosis, radiotherapy planning, preoperative navigation, and postoperative follow-up, forming a crucial foundation for intelligent medical image analysis. Compared to natural image segmentation, medical images exhibit significant differences. Medical images have complex grayscale distributions, low contrast, and high noise levels, and the scanning results vary considerably depending on the equipment or imaging conditions. Directly applying conventional segmentation algorithms often leads to a significant performance degradation. For abdominal multi-organ segmentation tasks, due to the large differences in organ morphology, blurred tissue boundaries, and imbalanced class proportions, it is difficult for models to simultaneously ensure accuracy for both the dominant and minority classes.

[0003] Currently, fully supervised segmentation methods based on deep learning (such as U-Net and its variants) can achieve good segmentation results when there is sufficient labeled data, but obtaining high-quality annotations is extremely difficult. Medical image annotation usually requires professional doctors to delineate the target region layer by layer in three-dimensional space, a process that is tedious, time-consuming, and costly. Therefore, labeled samples are often scarce in real-world scenarios, while unlabeled samples are relatively abundant, making semi-supervised medical image segmentation a research hotspot in recent years.

[0004] However, current state-of-the-art semi-supervised abdominal multi-organ medical image segmentation methods still face numerous technical challenges. In semi-supervised learning scenarios, labeled and unlabeled data typically originate from different hospitals, scanning equipment, or imaging parameters, leading to significant differences in feature distribution and a noticeable distribution shift. This causes systematic biases in the model's predictions on unlabeled samples. Simultaneously, abdominal multi-organ medical image data commonly suffers from class imbalance, with smaller organs (such as adrenal glands and veins) having a much lower pixel ratio than major organs. This makes the model more likely to focus on features of large organs during training, neglecting minority classes and resulting in decreased prediction accuracy. More complexly, distribution shift and class imbalance are often intertwined: distribution shift degrades the quality of pseudo-labels for minority classes, while class imbalance makes it difficult for the model to learn stable features from minority class samples, creating a vicious cycle of error accumulation. Furthermore, medical images themselves suffer from limited information density, noise interference, and blurred boundaries, requiring multi-dimensional feature modeling and cross-layer information fusion strategies to compensate for inherent imaging deficiencies. However, existing semi-supervised segmentation methods still have significant limitations in feature consistency modeling, class adaptive calibration, and pseudo-label quality optimization. They fail to fully explore the potential structural information in images, resulting in insufficient model generalization ability, especially in low-label sample scenarios where segmentation performance degrades more significantly. Summary of the Invention

[0005] To address the problems of degraded pseudo-label quality, difficulty in identifying a few organ classes, and uncontrollable generated samples in existing semi-supervised abdominal multi-organ segmentation methods under conditions of distributional shift between labeled and unlabeled data and extreme class imbalance, which lead to significant performance degradation in low-label scenarios, this invention provides a semi-supervised closed-loop optimization-based method for abdominal multi-organ medical image segmentation. This method achieves adaptive calibration of distributional shifts through a class-sensitive temperature correction strategy and constructs a closed-loop feedback mechanism for data generation and pseudo-label updating by combining a difference-aware diffusion sampling strategy. This allows for high-precision multi-organ segmentation results even with limited labeled data.

[0006] The technical means employed in this invention are as follows: A semi-supervised multi-organ image segmentation method to address class imbalance and distribution shift includes the following steps: S1. Obtain the abdominal multi-organ medical image dataset, divide the abdominal multi-organ medical image dataset into a training set and a test set. The training set includes a labeled dataset with real labels and an unlabeled dataset containing only images. Perform data augmentation operations on the images in the labeled dataset and the unlabeled dataset, and convert all images into three-dimensional voxel tensors of uniform size. S2. Construct a dual-head decoupled 3D neural network model. Use the labeled dataset tensor obtained in S1 as the supervision signal and the unlabeled dataset tensor as the consistency constraint for training. During training, dynamically calculate the temperature parameters of each category based on the distribution difference between labeled and unlabeled data. Perform class-sensitive temperature calibration on the predicted logical values ​​output by the network to obtain the calibrated supervised loss and unsupervised loss. Jointly optimize and update the network parameters, and output the trained dual-head network model and the corresponding category temperature parameters. S3. Apply the trained dual-head network model to the input tensor, decode the encoded features through the decoder of the trained dual-head network model, generate multi-organ segmentation prediction results, and calculate the segmentation confidence of each category. S4. Based on the segmentation confidence of each category, identify the difficult categories with low confidence; based on the difficult categories with low confidence, modify the real labels of the labeled dataset accordingly, and use the modified labels as conditions to input into the difference-aware diffusion model to generate synthetic samples that enhance the minority class. S5. Incorporate the enhanced minority class synthetic samples into the unlabeled dataset in S1 to form an updated semi-supervised dataset. Use the updated semi-supervised dataset to repeat S2 to S4 for training. When the model's segmentation performance on the validation set no longer improves in several consecutive training rounds, or when the number of training rounds reaches the preset maximum number of training rounds, end the training process and output the trained dual-head decoupled 3D neural network model. S6. Input the abdominal multi-organ medical images to be segmented into the trained dual-head decoupled 3D neural network model; perform forward inference on the input medical images through the trained dual-head decoupled 3D neural network model, and output the corresponding multi-organ segmentation results.

[0007] Furthermore, S1 specifically includes the following steps: The abdominal multi-organ medical image dataset is divided into a training set and a test set with a fixed number of images. The training set is further divided into a labeled dataset and an unlabeled dataset. The labeled dataset contains the input images and their corresponding ground truth labels, while the unlabeled dataset contains only the input images. Data augmentation was performed on the training set data. This included resizing the input abdominal multi-organ images to a D-dimensional size. The enhanced 3D images are converted into voxel tensors that can be processed by the network by random flipping and cropping with random probability, and then input into the training network in batches. The 3D images in the test set are sized and tensorized to generate corresponding 3D input tensors.

[0008] Furthermore, the voxel size of the three-dimensional voxel tensor is set to D. =64×128×128; The batch size input during training is set to 2.

[0009] Furthermore, S2 specifically includes the following steps: A three-dimensional convolutional feature extraction network consisting of a five-level downsampling encoder and a corresponding upsampling structure is constructed to encode the input data and form multi-scale features; Based on shared features, two independent first decoding branches are set. With the second decoding branch Output the predicted probabilities respectively and ; For labeled samples, the supervision loss is calculated using the first decoding branch and the second decoding branch respectively. and And a temperature-sensitive adjustment term is added to logits to achieve distributed calibration, in the following form:

[0010]

[0011] in, and For the two uncalibrated logits, For the category prior distribution, and The temperature coefficient is dynamically calculated based on category consistency and the degree of confusion. For unlabeled samples, the unsupervised loss is calculated as follows:

[0012]

[0013] in, and These are pseudo-labels generated based on the predicted output. The final joint optimization objective is: +

[0014] The parameters of the encoder and both decoding branches are updated simultaneously through backpropagation.

[0015] Furthermore, the weight coefficients in the joint optimization objective λ The value is 0.5.

[0016] Furthermore, the aforementioned sensitive temperature coefficient The steps to obtain it are as follows: Based on the correspondence between the true category labels in the labeled data and the pseudo labels generated by the dual-branch network on the unlabeled data, a category consistency-confusion matrix is ​​constructed; and the category consistency index and category confusion index are calculated for each category. Based on class consistency and class confusion indices, a local weight coefficient representing the difficulty of class segmentation is constructed. ; By statistically analyzing the overall consistency fluctuations and confusion distributions across all categories, a scaling factor q reflecting the global stability of the network is obtained. Finally, the sensitive temperature parameters were obtained. Used to calculate the temperature coefficients of the two supervisory branches. and .

[0017] Furthermore, S3 specifically includes the following steps: Spatial resolution is gradually restored by using multi-level upsampling, and skip connections are used in each layer to fuse shallow texture details and deep semantic information; The prediction results of the two branches, after being calibrated to be sensitive to temperature, are fused at the voxel level to generate the final 3D segmentation result.

[0018] Furthermore, S4 specifically includes the following steps: Based on the model's prediction results on the labeled set, the Dice index for each foreground category is calculated, and the k worst-performing categories are selected as the set of difficult categories. For the true label image of the labeled sample, extract the region mask belonging to the set of difficult categories and perform morphological dilation to enhance the structural saliency of the region, and obtain the modified label m; The modified label m is used as a condition and input into the difference-aware diffusion model to guide the generation of a 3D synthetic image containing structural information of the difficult category.

[0019] Furthermore, in S4, the method for determining the number of synthesized samples is as follows: Calculate the L2 distance between the class consistency vectors output by the dual-branch network as a measure of the dissimilarity d in the current training stability; Based on the difference degree d, the number of synthetic samples to be generated in this round is determined using a piecewise linear mapping function. The mapping rules are as follows:

[0020] Where D represents the maximum allowable range of differences; This indicates the maximum number of synthetic samples that can be generated in the current period. For intermediate sampling size, This is the minimum sampling size.

[0021] Furthermore, S5 specifically includes the following steps: The synthetic samples generated in the current training cycle are added to the unlabeled dataset and participate in the next round of training together with the original unlabeled samples. Before the start of a new training cycle, update the difficult category regions and sample generation strategies based on the prediction results of the previous round of model; By iteratively executing S2 to S4, a closed-loop process of dynamic data expansion and continuous model optimization is achieved.

[0022] Compared with the prior art, the present invention has the following advantages: This invention addresses the distribution shift and class imbalance issues between labeled and unlabeled data by constructing a unified optimization framework that integrates a class-sensitive temperature correction strategy and a difference-aware diffusion sampling strategy. First, a dual-head decoupling structure is used to model the confusion relationships between classes, and combined with global-local joint adjustment, dynamic class calibration is achieved, effectively mitigating prediction bias caused by distribution shift. Second, based on model uncertainty feedback, a difference-driven data synthesis mechanism is employed to adaptively enhance minority class samples, improving the model's ability to identify small organs and complex boundaries. Finally, the two strategies are fused into a closed-loop optimization system, achieving synergistic improvement in distribution adaptation and class balance. Experimental results on an abdominal multi-organ dataset demonstrate that this method significantly outperforms existing methods in both overall accuracy and small organ segmentation performance, exhibiting good generalization ability and clinical application value, and can be widely applied in the fields of automatic medical image segmentation and computer-aided diagnosis. Attached Figure Description

[0023] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the abdominal multi-organ CT data input by the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a framework diagram of the method of the present invention; Figure 4 This is a schematic diagram of the structure of the temperature parameter of the loss function of the present invention; Figure 5 This is a comparative experimental diagram of half-supervised medical image segmentation methods on an abdominal multi-organ dataset. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] like Figure 2 and Figure 3 As shown, this invention provides a semi-supervised multi-organ image segmentation method to solve class imbalance and distribution shift, comprising the following steps: S1. Obtain the abdominal multi-organ medical image dataset, divide the abdominal multi-organ medical image dataset into a training set and a test set. The training set includes a labeled dataset with real labels and an unlabeled dataset containing only images. Perform data augmentation operations on the images in the labeled dataset and the unlabeled dataset, and convert all images into three-dimensional voxel tensors of uniform size. S11. Obtain 3D images of multiple organs in the abdomen: The input is a 2D slice diagram of multiple abdominal organs. Figure 1 As shown, the abdominal multi-organ dataset is divided into a training set and a test set according to a fixed ratio; both the training set and the test set contain two sub-datasets, namely the input image and the corresponding ground truth label.

[0028] Based on the organizational structure of the abdominal multi-organ dataset, the training set is further divided into labeled and unlabeled datasets. The labeled dataset contains both the input image and its corresponding ground truth label, while the unlabeled dataset contains only the input image. During the training phase, both the labeled and unlabeled datasets participate in model training to achieve the semi-supervised closed-loop optimization proposed in this invention.

[0029] S12. Adjust the input abdominal 3D image to a uniform resolution. To best match the network structure.

[0030] In this embodiment, the voxel size is 64 × 128 × 128. Enhancement strategies such as random 3D cropping and random flipping are implemented with random scaling and probabilities to improve the model's robustness to spatial transformations and reduce overfitting. The enhanced 3D images are converted into a network-processable 3D voxel tensor format and packaged in batches using a data loader. In this embodiment, the batch size during the training phase is set to 2 to ensure a balance between training efficiency and memory usage.

[0031] S2. Construct a dual-head decoupled 3D neural network model. Use the labeled dataset tensor obtained in S1 as the supervision signal and the unlabeled dataset tensor as the consistency constraint for training. During training, dynamically calculate the temperature parameters of each category based on the distribution difference between labeled and unlabeled data. Perform class-sensitive temperature calibration on the predicted logical values ​​output by the network to obtain the calibrated supervised loss and unsupervised loss. Jointly optimize and update the network parameters, and output the trained dual-head network model and the corresponding category temperature parameters. S21. Feed the input tensor into the 3D convolutional encoder to obtain multi-scale features.

[0032] This step employs a three-dimensional convolutional feature extraction network consisting of a five-level downsampling encoder and a corresponding upsampling structure. This network takes the input tensor as input and extracts spatial semantic information at different scales through progressive downsampling and residual convolution operations. After processing by the feature encoder, five multi-scale coded features with different spatial resolutions and channel numbers are obtained, corresponding to feature representations from shallow to deep layers.

[0033] S22. Construct a dual-headed decoupled prediction branch and obtain complementary prediction outputs.

[0034] Based on shared coding features, two decoding branches with independent parameters are set. and To maintain structural decoupling and capture complementary information, the predicted probabilities are output separately. and By using parallel prediction outputs from two branches, the confusion patterns between categories can be characterized from two perspectives, providing an information basis for subsequent class-sensitive correction.

[0035] S23. Dynamically calculate the class-sensitive temperature coefficient based on class consistency and confusion level. .

[0036] In this invention and The calculation method is the same except for the row and column dimensions. For example Figure 4 The image shows one of them. The calculation method involves sorting and fusing consistency and confusion information according to category importance and error severity to construct local weight coefficients that characterize class-level difficulty. This allows for stronger modulation of categories that are difficult to distinguish or easily misclassified; simultaneously, by statistically analyzing the overall consistency fluctuations and confusion distributions of all categories, a scaling factor reflecting the global stability of the network can be obtained. This is used to adjust the overall correction force. The final result is the sensitive temperature parameter. .

[0037] S24. Calculate the supervised loss for labeled samples after temperature-sensitive adjustment.

[0038] For samples with real labels, the supervised loss is calculated using two branches. and And a temperature-sensitive adjustment term is added to logits to achieve distributed calibration, in the form of:

[0039]

[0040] S25. Construct the joint optimization loss function. The final loss function of this invention is: +

[0041] in, To monitor losses; This is an unsupervised loss. =0.5.

[0042] S3. Apply the trained dual-head network model to the input tensor, decode the encoded features through the decoder of the trained dual-head network model, generate multi-organ segmentation prediction results, and calculate the segmentation confidence of each category. S4. Based on the segmentation confidence of each category, identify the difficult categories with low confidence; based on the difficult categories with low confidence, modify the real labels of the labeled dataset accordingly, and use the modified labels as conditions to input into the difference-aware diffusion model to generate synthetic samples that enhance the minority class. S41. Construct a set of difficult categories based on the prediction performance of the dual-head decoupling network.

[0043] By using the prediction results of the dual-head decoupled network on unlabeled data, the Dice index of each foreground category on labeled data is calculated, and the k worst-performing categories are selected as the set of weak-performing categories to form a set of difficult categories, which is used to identify the categories most likely to be misclassified or missed in the current training stage.

[0044] This set is used for subsequent structure amplification and enhancement generation processes, enabling the generative model to specifically improve the expressive power of minority and weak classes. In this embodiment, k = 5.

[0045] S42. Perform structural saliency amplification on difficult category regions.

[0046] For each labeled sample, the regions belonging to the hard category set in its ground truth label image are first extracted using a masking method. Then, a morphological dilation operation is performed on these regions to increase the proportion of structural voxels, enhance category boundary features, and improve region coherence. The processed label image is denoted as m, and it serves as the conditional input for the subsequent diffusion model to strengthen the guidance capability of hard categories, thereby giving the generated image a more defined structural morphology.

[0047] S43. Determine the number of generated samples based on piecewise linear mapping of consistency difference.

[0048] To ensure that the number of generated samples adaptively matches the stability of the model in the current training phase, this embodiment determines the number of synthetic samples to be generated in this round through the following steps: First, calculate the L2 distance of the bi-branch class consistency vector to obtain the dissimilarity degree d, which is used to measure the stability of the network prediction for each class; Second, determine the number of synthetic samples based on the interval of the dissimilarity degree d through a piecewise function.

[0049] S44. Generate class-balanced 3D synthetic samples using a frozen conditional diffusion model. Input the modified label *m* into the pre-trained conditional diffusion model with its parameters frozen in this stage. Through the constraint effect of *m* in the reverse denoising process, guide the model to generate 3D organizational structures related to the morphology of difficult categories. The generated samples are then quality-selected based on the realism, consistency, and difficulty characteristics of their category structures, forming the synthetic sample set for the current training period. This sample set is subsequently used to expand the unlabeled data pool, providing a new data foundation for the next stage.

[0050] S5. Incorporate the enhanced minority class synthetic samples into the unlabeled dataset in S1 to form an updated semi-supervised dataset. Use the updated semi-supervised dataset to repeat S2 to S4 for training. When the model's segmentation performance on the validation set no longer improves in several consecutive training rounds, or when the number of training rounds reaches the preset maximum number of training rounds, end the training process and output the trained dual-head decoupled 3D neural network model. After the current training cycle ends, the latest synthetic samples generated by the conditional diffusion model are added to the unlabeled dataset and used together with the original unlabeled samples as input data for the next round of training, thereby continuously expanding the data coverage that the model can learn. In this way, the network can access augmented data generated based on the latest class difference information in each training round, realizing a data augmentation process that updates in real time as training progresses. Ultimately, this achieves stable, class-sensitive semi-supervised optimization results based on closed-loop feedback.

[0051] S6. Input the abdominal multi-organ medical images to be segmented into the trained dual-head decoupled 3D neural network model; perform forward inference on the input medical images through the trained dual-head decoupled 3D neural network model, and output the corresponding multi-organ segmentation results.

[0052] Figure 5 This figure shows a comparative experiment of semi-supervised medical image segmentation methods on an abdominal multi-organ dataset. The figure displays two sets of segmentation results: the first column shows the ground truth labels, the second column shows the segmentation results of the method described in this invention, and the third, fourth, fifth, and sixth columns show the segmentation results of other methods. Figure 5 As can be seen, compared with the comparison method, the method of the present invention can effectively segment a few organ classes, such as the gallbladder highlighted in blue, which suffers from severe missed detection or even inability to be segmented in the segmentation results of the comparison method. Furthermore, the second set of segmentation results shows that the method of the present invention has a significant advantage in mitigating class confusion: the comparison method is prone to large-scale misclassification between foreground and background, and exhibits obvious boundary blurring problems in adjacent organ regions; in contrast, the method of the present invention can obtain cleaner background segmentation results and more accurately segment the boundaries of adjacent organs, thereby improving the overall accuracy and stability of multi-organ segmentation.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semi-supervised multi-organ image segmentation method to solve class imbalance and distribution shift, characterized in that, Includes the following steps: S1. Obtain the abdominal multi-organ medical image dataset and divide it into a training set and a test set. The training set includes a labeled dataset with real labels and an unlabeled dataset containing only images. Data augmentation operations are performed on images in both labeled and unlabeled datasets, and all images are converted into three-dimensional voxel tensors of uniform size. S2. Construct a dual-head decoupled three-dimensional neural network model, using the labeled dataset tensor obtained in S1 as the supervision signal and the unlabeled dataset tensor as the consistency constraint for training; During training, the temperature parameters of each category are dynamically calculated based on the distribution differences between labeled and unlabeled data. The predicted logical values ​​output by the network are calibrated to be sensitive to temperature, and the calibrated supervised loss and unsupervised loss are obtained. The network parameters are jointly optimized and updated, and the trained dual-head network model and the corresponding category temperature parameters are output. S3. Apply the trained dual-head network model to the input tensor, decode the encoded features through the decoder of the trained dual-head network model, generate multi-organ segmentation prediction results, and calculate the segmentation confidence of each category. S4. Based on the segmentation confidence of each category, identify the difficult categories with low confidence; based on the difficult categories with low confidence, modify the real labels of the labeled dataset accordingly, and use the modified labels as conditions to input into the difference-aware diffusion model to generate synthetic samples that enhance the minority class. S5. Incorporate the enhanced minority class synthetic samples into the unlabeled dataset in S1 to form an updated semi-supervised dataset. Use the updated semi-supervised dataset to repeat S2 to S4 for training. When the model's segmentation performance on the validation set no longer improves in several consecutive training rounds, or when the number of training rounds reaches the preset maximum number of training rounds, end the training process and output the trained dual-head decoupled 3D neural network model. S6. Input the abdominal multi-organ medical images to be segmented into the trained dual-head decoupled 3D neural network model; perform forward inference on the input medical images through the trained dual-head decoupled 3D neural network model, and output the corresponding multi-organ segmentation results.

2. The semi-supervised multi-organ image segmentation method for resolving class imbalance and distribution shift according to claim 1, characterized in that, S1 specifically includes the following steps: The abdominal multi-organ medical image dataset is divided into a training set and a test set with a fixed number of images. The training set is further divided into a labeled dataset and an unlabeled dataset. The labeled dataset contains the input images and their corresponding ground truth labels, while the unlabeled dataset contains only the input images. Data augmentation was performed on the training set data. This included resizing the input abdominal multi-organ images to a D-dimensional size. The enhanced 3D images are converted into voxel tensors that can be processed by the network by random flipping and cropping with random probability, and then input into the training network in batches. The 3D images in the test set are sized and tensorized to generate corresponding 3D input tensors.

3. The semi-supervised multi-organ image segmentation method for resolving class imbalance and distribution shift according to claim 2, characterized in that, The voxel size of the three-dimensional voxel tensor is set to D. =64×128×128; The batch size input during training is set to 2.

4. The semi-supervised multi-organ image segmentation method for resolving class imbalance and distribution shift according to claim 1, characterized in that, S2 specifically includes the following steps: A three-dimensional convolutional feature extraction network consisting of a five-level downsampling encoder and a corresponding upsampling structure is constructed to encode the input data and form multi-scale features; Based on shared features, two independent first decoding branches are set. With the second decoding branch Output the predicted probabilities respectively and ; For labeled samples, the supervision loss is calculated using the first decoding branch and the second decoding branch respectively. and And a temperature-sensitive adjustment term is added to logits to achieve distributed calibration, in the following form: in, and For the two uncalibrated logits, For the category prior distribution, and The temperature coefficient is dynamically calculated based on category consistency and the degree of confusion. For unlabeled samples, the unsupervised loss is calculated as follows: in, and These are pseudo-labels generated based on the predicted output. The final joint optimization objective is: + The parameters of the encoder and both decoding branches are updated simultaneously through backpropagation.

5. The semi-supervised multi-organ image segmentation method for resolving class imbalance and distribution shift according to claim 4, characterized in that, The weight coefficients in the joint optimization objective λ The value is 0.

5.

6. The semi-supervised multi-organ image segmentation method for resolving class imbalance and distribution shift according to claim 4, characterized in that, The type of sensitive temperature coefficient The steps to obtain it are as follows: Based on the correspondence between the true category labels in the labeled data and the pseudo labels generated by the dual-branch network on the unlabeled data, a category consistency-confusion matrix is ​​constructed; and the category consistency index and category confusion index are calculated for each category. Based on class consistency and class confusion indices, a local weight coefficient representing the difficulty of class segmentation is constructed. ; By statistically analyzing the overall consistency fluctuations and confusion distributions across all categories, a scaling factor q reflecting the global stability of the network is obtained. Finally, the sensitive temperature parameters were obtained. Used to calculate the temperature coefficients of the two supervisory branches. and .

7. The semi-supervised multi-organ image segmentation method for resolving class imbalance and distribution shift according to claim 1, characterized in that, S3 specifically includes the following steps: Spatial resolution is gradually restored by using multi-level upsampling, and skip connections are used in each layer to fuse shallow texture details and deep semantic information; The prediction results of the two branches, after being calibrated to be sensitive to temperature, are fused at the voxel level to generate the final 3D segmentation result.

8. The semi-supervised multi-organ image segmentation method for resolving class imbalance and distribution shift according to claim 1, characterized in that, S4 specifically includes the following steps: Based on the model's prediction results on the labeled set, the Dice index for each foreground category is calculated, and the k worst-performing categories are selected as the set of difficult categories. For the true label image of the labeled sample, extract the region mask belonging to the set of difficult categories and perform morphological dilation to enhance the structural saliency of the region, and obtain the modified label m; The modified label m is used as a condition and input into the difference-aware diffusion model to guide the generation of a 3D synthetic image containing structural information of the difficult category.

9. The semi-supervised multi-organ image segmentation method for resolving class imbalance and distribution shift according to claim 8, characterized in that, In S4, the method for determining the number of synthetic samples generated is as follows: Calculate the L2 distance between the class consistency vectors output by the dual-branch network as a measure of the dissimilarity d in the current training stability; Based on the difference degree d, the number of synthetic samples to be generated in this round is determined using a piecewise linear mapping function. The mapping rules are as follows: Where D represents the maximum allowable range of differences; This indicates the maximum number of synthetic samples that can be generated in the current period. For intermediate sampling size, This is the minimum sampling size.

10. The semi-supervised multi-organ image segmentation method for resolving class imbalance and distribution shift according to claim 1, characterized in that, S5 specifically includes the following steps: The synthetic samples generated in the current training cycle are added to the unlabeled dataset and participate in the next round of training together with the original unlabeled samples. Before the start of a new training cycle, update the difficult category regions and sample generation strategies based on the prediction results of the previous round of model; By iteratively executing S2 to S4, a closed-loop process of dynamic data expansion and continuous model optimization is achieved.

Images