CN-121981982-A - Semi-supervised medical image segmentation method based on dynamic ambiguity perception

CN121981982ACN 121981982 ACN121981982 ACN 121981982ACN-121981982-A

Abstract

The invention discloses a semi-supervised medical image segmentation method based on dynamic ambiguity sensing, which designs a dynamic ambiguity sensing framework, comprising a heterogeneous feature extraction module, a dynamic ambiguity bootstrapping module and an ambiguity sensing module. Firstly, an original medical image enters a heterogeneous feature extraction module, and two competing subnets extract high-level semantic features and generate different pseudo tags. The pseudo labels generated by the two competing subnetworks enter a dynamic ambiguity bootstrapping module, and the competing subnetworks are divided into a strong network and a weak network by calculating the quality scores of the two pseudo labels. After the strong network and the weak network are divided, the prediction generated by the weak network focuses on the mutual ambiguity perception of the targeted learning of the pseudo tag knowledge, and the prediction generated by the strong network focuses on the self-ambiguity perception of the self-ambiguity examination. The invention improves the accuracy of medical image segmentation, can effectively reduce the diagnosis time of doctors and improves the efficiency. Meanwhile, missing diagnosis or misdiagnosis caused by different diagnosis levels or fatigue of doctors is reduced.

Inventors

SUN NING
Tang Daodao
SUN SIPEI

Assignees

南京邮电大学

Dates

Publication Date: 20260505
Application Date: 20260119

Claims (9)

1. A semi-supervised medical image segmentation method based on dynamic ambiguity perception is characterized by comprising the following steps: s1, acquiring a medical image dataset, wherein the dataset comprises marked medical images and unmarked medical images; S2, constructing a semi-supervised medical image segmentation model based on dynamic ambiguity perception, wherein the semi-supervised medical image segmentation model comprises a heterogeneous feature extraction module, a dynamic ambiguity bootstrapping module and an ambiguity perception module, and the semi-supervised medical image segmentation model is used for processing medical images, and specifically comprises the following steps: Inputting the medical image into the heterogeneous feature extraction module, and respectively extracting features by using two competing subnets with different structures and generating pseudo tags; Inputting the pseudo tag into a dynamic ambiguity bootstrapping module, calculating the quality score of the pseudo tag generated by each competition sub-network, and dynamically dividing the competition sub-network into a strong network and a weak network according to the quality score, wherein the quality of the pseudo tag generated by the strong network is higher than that of the pseudo tag generated by the weak network; Finally, inputting the pseudo tag into an ambiguity sensing module, wherein the ambiguity sensing module comprises a mutual ambiguity sensing unit and a self-ambiguity sensing unit, wherein in the mutual ambiguity sensing unit, the weak network learns the pseudo tag knowledge of the strong network in an ambiguity area; s3, training the semi-supervised medical image segmentation model based on the supervision loss of the annotated medical image and the self-supervision loss of the unlabeled medical image generated by pseudo-tag and ambiguity perception; S4, predicting the medical image to be segmented by using the trained semi-supervised medical image segmentation model, and generating a segmentation result.
2. The method of claim 1, wherein the medical image dataset of step S1 includes a large number of unlabeled brain CTA scan images, left atrium MRI scan images, and pancreatic CT scan images, and a small number or more of highly accurately labeled scan images of the three medical images.
3. The method for segmenting the semi-supervised medical image based on dynamic ambiguity perception according to claim 1, wherein in the step S2, the medical image is input into the heterogeneous feature extraction module, features are respectively extracted by using two competing subnets with different structures and pseudo labels are generated, and the method specifically comprises the following steps: two deep learning networks with different structures are used as two parallel competition subnets, and a heterogeneous feature extraction module is constructed; The medical image is respectively input into two competition subnets, the two competition subnets respectively extract high-level semantic features of the medical image, and the extracted features respectively generate pseudo tags of the two competition subnets after cross entropy.
4. The method of claim 1, wherein the step S2 is characterized in that the step of inputting the pseudo tag into the dynamic ambiguity bootstrapping module calculates a quality score of the pseudo tag generated by each competing subnet and dynamically divides the competing subnets into a strong network and a weak network according to the quality score, and the method specifically comprises the steps of: Voxel location for pseudo-labels The prediction probability is recorded as Indicating that the voxel position belongs to the category Predictive probability of (1), wherein , , , Respectively representing the height, width and depth dimensions of the image, and the category number ; Voxel location for pseudo-labels The ambiguity of voxel prediction at this point is noted as: ; the mass fraction of the pseudo tag is calculated, and the calculation formula is as follows: ; And comparing the quality scores of the pseudo labels obtained by the two competing subnets, determining the competing subnets with high quality scores as strong networks, and positioning the competing subnets with low quality scores as weak networks.
5. The method for semi-supervised medical image segmentation based on dynamic ambiguity perception as set forth in claim 4, wherein in the mutual ambiguity perception unit in step S2, the weak network learns pseudo tag knowledge of the strong network in an ambiguity region, specifically: setting an ambiguity threshold T, and generating an auxiliary ambiguity graph based on a comparison result of the voxel ambiguity predicted by the weak network and a preset threshold T, wherein the auxiliary ambiguity graph is expressed as: ; then generating a mutual ambiguity graph based on the prediction difference of the strong network and the weak network, which is expressed as follows: ; multiplying the auxiliary ambiguity graph with the mutual ambiguity graph to obtain a final ambiguity graph, wherein the final ambiguity graph is expressed as: ; And based on the final ambiguity graph, calculating the loss, and determining the area and the weight of the weak network to learn to the strong network.
6. The method according to claim 5, wherein in the self-ambiguity sensing unit in step S2, the input of the strong network is disturbed and self-ambiguity sensing is performed based on prediction consistency, and the method specifically comprises: Applying slight disturbance to the original image input into the strong network to obtain a disturbance image, and inputting the disturbance image into the strong network to obtain a disturbance prediction result; And constructing a self-ambiguity joint confidence matrix based on the prediction of the strong network on the original image and the disturbance prediction result, then further solving a self-ambiguity joint distribution matrix, identifying unstable voxels in the strong network prediction, and carrying out consistency regularization constraint on the prediction of the unstable voxels.
7. The semi-supervised medical image segmentation method based on dynamic ambiguity perception as set forth in claim 6, wherein the construction of the self-ambiguity joint confidence matrix C is as follows: ; the construction of the self-ambiguity joint distribution matrix Q is as follows: ; Wherein, the The image to be input is represented by a representation, Representing voxels; represented as a noisy prediction of a strong network, Representing voxels Belonging to Is a function of the probability of (1), As a potential for a true tag to be present, Represented as The number of voxels; the number of categories is indicated and, Representing the average confidence.
8. The semi-supervised medical image segmentation method based on dynamic ambiguity perception according to claim 6, wherein said supervised loss in step S3 employs a cross entropy loss function, said self-supervised loss comprising a mutual ambiguity perceived loss weighted based on said final ambiguity map and a self-ambiguity perceived loss calculated based on the consistency deviation of unstable voxels.
9. The method for segmenting a semi-supervised medical image based on dynamic ambiguity perception according to claim 1, wherein in step S4, the labeled medical image is used for testing before the medical image to be segmented is predicted, specifically: performing medical image segmentation after using the weight average of the two trained competition sub-network models; And verifying the accuracy and model performance of the segmentation result based on the segmentation result and the labeling, including: The accuracy of the whole segmentation is calculated by the proportion of the correctly predicted voxels to the total voxels, and the segmentation intersection ratio is calculated by comparing the ratio of the correctly predicted number of the voxels to the union of the true number of the entities in the medical image and the prediction result, so as to evaluate the model performance.

Description

Semi-supervised medical image segmentation method based on dynamic ambiguity perception Technical Field The invention belongs to the technical field of three-dimensional image processing, and particularly relates to a semi-supervised medical image segmentation method based on dynamic ambiguity perception. Background Deep learning has shown significant potential in the field of medical image segmentation, and its diagnostic accuracy can even exceed the level of human expert, especially in the task of identifying complex structures such as tumors, organs, etc. However, the implementation of high-performance models typically relies on large-scale high-quality labeling data, while the labeling process of medical images requires deep involvement of a specialist, resulting in high costs and long cycles. With the rapid popularization of medical imaging technology, the amount of unlabeled image data grows exponentially, and how to efficiently utilize limited labeled data and massive unlabeled data becomes a core bottleneck restricting technology landing. Currently, semi-supervised learning is considered as a key approach to solve this contradiction. The method assists in improving the generalization capability of the model by mining potential rules in unlabeled data. However, the existing methods face many challenges in the practical process, firstly, noise is easily introduced into a pseudo tag generation mechanism, errors are continuously accumulated and propagated along with training iteration, and finally model performance is degraded, secondly, most of the methods adopt fixed teacher-student network architecture, lack dynamic adjustment capability and are difficult to adapt to learning requirements of pseudo tags with different qualities, and in addition, the traditional methods focus on consistency constraint of tag space, and have insufficient attention to separability of feature space, so that capturing capability of the model on a microstructure is limited. Notably, the target region in the medical image often has the characteristics of changeable morphology and small duty ratio (such as angiopathy and micro tumor), and the voxel level average strategy commonly used in the prior method is extremely easy to lose key detail information. Although recent studies have attempted to alleviate the above problems by uncertainty quantization or multi-model collaboration, there are drawbacks of insufficient information utilization, rough error suppression mechanism, and the like. Therefore, developing a semi-supervised learning framework capable of dynamically sensing fuzzy areas, adaptively screening high-quality knowledge and realizing multi-dimensional constraint becomes an important direction for breaking through the bottleneck of the current technology. Disclosure of Invention The present invention aims to solve at least one of the technical problems existing in the related art to some extent. The invention aims to provide a semi-supervised medical image segmentation method based on dynamic ambiguity perception, which improves the accuracy and the robustness of medical image segmentation. In order to achieve the above purpose, the invention provides a semi-supervised medical image segmentation method based on dynamic ambiguity perception, which comprises the following steps: s1, acquiring a medical image dataset, wherein the dataset comprises marked medical images and unmarked medical images; S2, constructing a semi-supervised medical image segmentation model based on dynamic ambiguity perception, wherein the semi-supervised medical image segmentation model comprises a heterogeneous feature extraction module, a dynamic ambiguity bootstrapping module and an ambiguity perception module, and the semi-supervised medical image segmentation model is used for processing medical images, and specifically comprises the following steps: Inputting the medical image into the heterogeneous feature extraction module, and respectively extracting features by using two competing subnets with different structures and generating pseudo tags; Inputting the pseudo tag into a dynamic ambiguity bootstrapping module, calculating the quality score of the pseudo tag generated by each competition sub-network, and dynamically dividing the competition sub-network into a strong network and a weak network according to the quality score, wherein the quality of the pseudo tag generated by the strong network is higher than that of the pseudo tag generated by the weak network; Finally, inputting the pseudo tag into an ambiguity sensing module, wherein the ambiguity sensing module comprises a mutual ambiguity sensing unit and a self-ambiguity sensing unit, wherein in the mutual ambiguity sensing unit, the weak network learns the pseudo tag knowledge of the strong network in an ambiguity area; s3, training the semi-supervised medical image segmentation model based on the supervision loss of the annotated medical image and the self-supervision loss o