CN-122023791-A - Semi-supervised medical image segmentation method based on asymmetric deformation guiding mutual learning

CN122023791ACN 122023791 ACN122023791 ACN 122023791ACN-122023791-A

Abstract

The invention discloses a semi-supervised medical image segmentation method and device based on asymmetric deformation guiding mutual learning, and relates to the technical field of image segmentation. The method comprises the steps of constructing an asymmetric double-branch mutual learning network architecture, dividing and preprocessing a data set to obtain marked data and unmarked data, inputting the marked data into the two networks, calculating supervision losses of the two networks, extracting an offset field through the deformed network to generate a deformed guide weight mask, constructing a standard network consistency loss, constructing a deformed network consistency loss according to a pseudo tag of the standard network, maintaining two branch feature libraries, calculating cross-model feature distribution alignment loss, fusing the supervision losses, the consistency loss and the feature alignment loss, updating parameters of the two networks to obtain a trained network, inputting an image to be segmented into the trained network, and outputting a final segmentation result. The embodiment of the invention can improve the segmentation precision of the medical image.

Inventors

KONG QINGQUN
PENG YU
ZOU HAOYU
YANG KEHU

Assignees

中国矿业大学(北京)

Dates

Publication Date: 20260512
Application Date: 20260109

Claims (10)

1. A semi-supervised medical image segmentation method based on asymmetric deformation guided mutual learning, the method comprising: S1, acquiring a medical image data set, dividing the medical image data set into a labeled data set and an unlabeled data set, preprocessing the medical image data set, and constructing a data loader; S2, constructing an asymmetric double-branch learning network architecture, wherein the architecture comprises two parallel segmentation networks, namely a standard segmentation network and a deformation segmentation network, respectively, inputting labeled sample data into the standard segmentation network and the deformation segmentation network respectively, and calculating the supervision loss of the standard segmentation network and the supervision loss of the deformation segmentation network; S3, inputting the unlabeled sample data into a standard segmentation network, and outputting a standard branch prediction probability map and a pseudo-label, inputting the unlabeled sample data into a deformation segmentation network, generating a deformation branch pseudo-label and a prediction probability map, and simultaneously generating an offset field of the middle layer; s4, calculating the consistency loss of the standard segmentation network based on the deformation guide weight mask, the standard branch prediction probability map and the pseudo tag of the deformation branch, and calculating the consistency loss of the deformation segmentation network based on the deformation branch prediction probability map and the pseudo tag of the standard branch; S5, constructing two dynamic feature libraries, wherein the first dynamic feature library is used for storing high-confidence feature distribution on the history of the standard segmentation network, and the second dynamic feature library is used for storing high-confidence feature distribution on the history of the deformed segmentation network; S6, constructing a total loss function based on asymmetric deformation guiding consistency loss, standard segmentation network supervision loss, deformation segmentation network supervision loss and overall characteristic loss, and updating parameters of the standard segmentation network and the deformation segmentation network through a back propagation algorithm to obtain a trained standard segmentation network and a trained deformation segmentation network; S7, acquiring a medical image to be segmented, inputting the medical image to be segmented into a trained standard segmentation network and a deformed segmentation network, and outputting a final focus segmentation result.
2. The semi-supervised medical image segmentation method based on asymmetric deformation guided mutual learning as set forth in claim 1, wherein the standard segmentation network adopts a V-Net structure and comprises an input layer, an encoder, a decoder and an output layer, wherein a projection head module is added to a specific level of the decoder for mapping high-dimensional features into a low-dimensional space; The standard segmentation network is used for extracting global context characteristics and stable semantic information; The deformable segmentation network adopts a V-Net-DCN structure and comprises an input layer, an encoder, a decoder and an output layer, wherein a convolution block of a deformable convolution unit is integrated in the encoder, the deformable convolution unit internally comprises an offset generation network which is a conventional convolution layer and is used for receiving an input characteristic diagram and outputting an offset field with a channel number of a specific multiple, and a projection head module is added to a specific level of the decoder; The deformation segmentation network is used for extracting local geometric deformation characteristics through learning offset.
3. The semi-supervised medical image segmentation method based on asymmetric deformation guided mutual learning as set forth in claim 1, wherein the step of calculating a spatial liveness map based on the offset field of S3, the step of calculating a deformation guide weight mask based on the spatial liveness map comprises: calculating an L2 norm based on the offset vector of the offset field, obtaining the displacement distance of each sampling point, and averaging all sampling points to obtain the average deformation strength of the voxel position, wherein the average deformation strength is a space liveness graph; and converting the liveness into weight by utilizing an exponential decay function based on the sampling activity map to generate a deformation guide weight mask.
4. A semi-supervised medical image segmentation method based on asymmetric deformation guided mutual learning as set forth in claim 3, wherein the specific process of computing a spatial liveness graph is represented by the following equation (1): (1) Wherein, the Represent the first The deformable convolution blocks are in position Spatial activity values at; Representing the convolution kernel size; representing the total number of three-dimensional sampling points; Represent the first Learning offset vectors for the sampling points; representing euclidean norms; the specific process of generating the deformation guide weight mask is represented by the following formula (2): (2) Wherein, the A super parameter indicating a control attenuation rate; Representing unlabeled sample data At voxels A deformation guide weight mask at the position; representing a sampling activity map.
5. The semi-supervised medical image segmentation method based on asymmetric deformation guided mutual learning as set forth in claim 1, wherein the consistency loss of the standard segmentation network is represented by the following equation (3): (3) Wherein, the Consistency loss of standard split network; representing unlabeled sample data; a confidence mask representing the deformed segmented network; representing a cross entropy loss function; representing a standard segmentation network at voxels Prediction output at; representing a pseudo tag generated by the deformation partition network; representing the deformation guide weight mask.
6. The semi-supervised medical image segmentation method based on asymmetric deformation guide mutual learning according to claim 1 is characterized in that the dynamic feature library is a fixed-length queue, wherein in each iteration, the proportion of high confidence pixels is calculated based on a predictive probability map of the obtained unlabeled sample data, and when the proportion of the high confidence pixels exceeds a preset threshold, features of the corresponding unlabeled sample data after being processed by a projection head are added into the dynamic feature library.
7. The semi-supervised medical image segmentation method based on asymmetric deformation guided mutual learning according to claim 1, wherein the step S5 of calculating the maximum mean difference between the feature distributions extracted by the two segmentation networks in the current iteration based on the two dynamic feature libraries, and constructing the overall feature loss comprises the following steps: Based on the two dynamic feature libraries, for a standard segmentation network, sampling a historical feature set from the dynamic feature library corresponding to the deformed segmentation network, calculating the maximum mean value difference between the historical feature set and the current batch feature set in the standard segmentation network, and constructing the feature loss of the standard segmentation network in a minimized mode; based on the two dynamic feature libraries, for the deformation segmentation network, sampling a historical feature set from the dynamic feature library corresponding to the standard segmentation network, calculating the maximum mean value difference between the historical feature set and the current batch feature set in the deformation segmentation network, and constructing the feature loss of the deformation segmentation network in a minimized mode; The overall feature loss is constructed based on the feature loss of the standard segmentation network and the feature loss of the deformed segmentation network.
8. Semi-supervised medical image segmentation apparatus based on asymmetric deformation guided mutual learning for implementing the semi-supervised medical image segmentation method based on asymmetric deformation guided mutual learning as set forth in any one of claims 1-7, the apparatus comprising: The system comprises a data acquisition unit, a data loader, a data processing unit and a data processing unit, wherein the data acquisition unit is used for acquiring a medical image data set and dividing the medical image data set into a labeled data set and an unlabeled data set; The system comprises a construction and calculation unit, a calculation unit and a calculation unit, wherein the construction and calculation unit is used for constructing an asymmetric double-branch learning network architecture, and the architecture comprises two parallel division networks which are a standard division network and a deformation division network respectively; The deformation guiding calculation unit is used for inputting the unlabeled sample data into the standard segmentation network and outputting a standard branch prediction probability map and a pseudo label, inputting the unlabeled sample data into the deformation segmentation network, generating a deformed branch pseudo label and a prediction probability map, and simultaneously generating an offset field of the middle layer; The system comprises a first construction unit, a second construction unit, a third construction unit, a fourth construction unit, a fifth construction unit and a sixth construction unit, wherein the first construction unit is used for calculating the consistency loss of the standard segmentation network based on the deformation guiding weight mask, the standard branch prediction probability map and the pseudo tag; The feature library management unit is used for constructing two dynamic feature libraries; the system comprises a first dynamic feature library, a second dynamic feature library, a third dynamic feature library, a fourth dynamic feature library, a fifth dynamic feature library and a fourth dynamic feature library, wherein the first dynamic feature library is used for storing high-confidence feature distribution on a standard segmentation network history; The training unit is used for constructing a total loss function based on asymmetric deformation guiding consistency loss, standard segmentation network supervision loss, deformation segmentation network supervision loss and overall characteristic loss, updating parameters of the standard segmentation network and the deformation segmentation network through a back propagation algorithm, and obtaining a trained standard segmentation network and a trained deformation segmentation network; The medical image to be segmented is input into a trained standard segmentation network and a deformation segmentation network, and a final focus segmentation result is output.
9. A semi-supervised medical image segmentation apparatus based on asymmetric deformation-guided mutual learning, characterized in that the semi-supervised medical image segmentation apparatus based on asymmetric deformation-guided mutual learning comprises: A processor; a memory having stored thereon computer readable instructions which, when executed by the processor, implement the method of any of claims 1 to 7.
10. A computer readable storage medium having stored therein program code which is callable by a processor to perform the method of any one of claims 1 to 7.

Description

Semi-supervised medical image segmentation method based on asymmetric deformation guiding mutual learning Technical Field The invention relates to the technical field of image segmentation, in particular to a semi-supervised medical image segmentation method and device based on asymmetric deformation guiding mutual learning. Background Medical image segmentation is a key link in medical image analysis, aims to accurately outline the outline of an organ or a focus from image data, and provides quantitative basis for clinical diagnosis, treatment planning and prognosis evaluation. With the development of deep learning technology, the full-supervised learning method achieves remarkable effects on medical image segmentation tasks. However, these models typically rely on high quality data for a large number of pixel level labels. In the medical field, the cost of acquiring the labeling data is extremely high, so that not only is a great deal of time spent by a deep radiologist required, but also the problem of data privacy is faced, and the labeling data is seriously scarcity. Semi-supervised learning has become a current research hotspot in order to reduce reliance on annotation data. Semi-supervised learning improves model performance by utilizing a small amount of annotated data and a large amount of unlabeled data simultaneously. Currently, the mainstream semi-supervised medical image segmentation method mostly adopts a collaborative training or mutual learning framework. These methods typically build two networks that are identical in structure, with the two networks supervising each other through consistency constraints. Although the existing homogeneous mutual learning method improves the segmentation performance to a certain extent, significant limitations still exist. First, existing co-training frameworks mostly employ two identical network architectures and the same parameter initialization. This means that both networks have the same generalized bias, tend to focus on similar image features, and produce similar mispredictions on unlabeled data. This homogeneity can lead to inter-reinforcement errors between networks, creating validation bias, limiting the upper limit of the model. Second, some current consistency learning methods typically assign the same weight to flat areas and complex boundary areas when computing a consistency loss, as a function of the consistency kernel for all pixels in the image. However, medical images often contain complex non-rigid deformations and blurred boundaries. For these geometrically complex regions, the uncertainty of the network predictions is high, direct forced consistency may introduce noise, while for structurally stable regions the network predictions are more reliable. The existing method lacks an effective mechanism to quantify the complexity of local geometric deformation, and the learning weight cannot be dynamically adjusted according to the complexity. Finally, most existing methods apply consistency constraints only at the network output layer, ignoring the rich semantic information contained in the intermediate feature layer. While some approaches have attempted feature alignment, it is often difficult to preserve category separability while guaranteeing consistency of feature distribution. In summary, the current medical image segmentation algorithm has the problem of network homogeneity, which results in poor accuracy and low accuracy of the final segmentation result. Disclosure of Invention In order to solve the technical problems of poor precision and low accuracy of a final segmentation result caused by network homogeneity in the prior art, the embodiment of the invention provides a semi-supervised medical image segmentation method and device based on asymmetric deformation guiding mutual learning. The technical scheme is as follows: in one aspect, a semi-supervised medical image segmentation method based on asymmetric deformation-guided mutual learning is provided, the method is implemented by a semi-supervised medical image segmentation device based on asymmetric deformation-guided mutual learning, and the method comprises: S1, acquiring a medical image data set, dividing the medical image data set into a labeled data set and an unlabeled data set, preprocessing the medical image data set, and constructing a data loader; S2, constructing an asymmetric double-branch learning network architecture, wherein the architecture comprises two parallel segmentation networks, namely a standard segmentation network and a deformation segmentation network, respectively, inputting labeled sample data into the standard segmentation network and the deformation segmentation network respectively, and calculating the supervision loss of the standard segmentation network and the supervision loss of the deformation segmentation network; S3, inputting the unlabeled sample data into a standard segmentation network, and outputting a standard branch prediction probabil