CN-121685538-B - Lung vessel CT image segmentation method, system, terminal and storage medium based on semi-supervised learning

CN121685538BCN 121685538 BCN121685538 BCN 121685538BCN-121685538-B

Abstract

The invention relates to the technical field of image processing and discloses a lung vessel CT image segmentation method, a system, a terminal and a storage medium based on semi-supervised learning, wherein the method comprises the steps of constructing a depth image segmentation model, wherein the model comprises an improved encoder, an improved jump connection layer and a decoder; the bottom layer of the improved encoder is provided with a trunk enhancement module, the first four downsampling stages are respectively provided with a blood vessel feature enhancement module, after the feature images output by each stage of the improved encoder are enhanced by the blood vessel feature enhancement module, the feature images based on the corresponding stages of the improved jump connection layer and the decoder are overlapped, the segmentation model is trained by adopting a multi-dimensional composite loss function based on a semi-supervision training framework to obtain a target image segmentation model, a lung blood vessel CT image to be processed is acquired and is input into the target image segmentation model for segmentation, and a segmentation result of the lung blood vessel CT image is output. The invention improves the accuracy of the CT image segmentation of the pulmonary blood vessels.

Inventors

LIANG ZHIQING
HUANG BINGDING
LI RUI
ZENG KE
SU LIYILEI

Assignees

深圳技术大学

Dates

Publication Date: 20260505
Application Date: 20260210

Claims (9)

1. The lung vessel CT image segmentation method based on semi-supervised learning is characterized by comprising the following steps of: Constructing a depth image segmentation model based on a U-shaped network architecture, wherein the depth image segmentation model comprises an improved encoder, an improved jump connection layer and a decoder; training the depth image segmentation model by adopting a multi-dimensional composite loss function based on a semi-supervised training framework to obtain a target image segmentation model; Acquiring a pulmonary vessel CT image to be processed, inputting the pulmonary vessel CT image to the improved encoder of the target image segmentation model for cascade downsampling and feature enhancement to obtain an enhancement feature image of each stage, and respectively carrying out residual superposition on the enhancement feature image of each stage and an upsampling feature image of a corresponding stage of the decoder through the improved jump connection layer until the upsampling of the decoder is completed, and outputting a segmentation result of the pulmonary vessel CT image; The improved encoder for inputting the pulmonary vessel CT image into the target image segmentation model carries out cascade downsampling and feature enhancement to obtain an enhancement feature map of each stage, which comprises the following steps: inputting the pulmonary vessel CT image to the improved encoder of the target image segmentation model, sequentially passing through five cascade downsampling stages; Each of the first four downsampling stages is processed through a residual convolution block and a blood vessel characteristic enhancement module in sequence, and the fifth downsampling stage is processed through a residual convolution block and a multi-scale time convolution 3D module in sequence; after the processing of each downsampling stage is finished, outputting an enhancement feature map of the current stage, and after all five downsampling stages are finished, enhancing and purifying the obtained deepest-level feature map through a trunk enhancement module to obtain a bottom enhancement feature map; And respectively transmitting the bottom enhancement feature map and enhancement feature maps output by the first four downsampling stages to the improved jump connection layer.
2. The pulmonary vessel CT image segmentation method based on semi-supervised learning of claim 1, wherein the training the depth image segmentation model with a multi-dimensional composite loss function based on a semi-supervised training framework to obtain a target image segmentation model further comprises: Acquiring original medical image scanning data stored in an NIFTI format and corresponding original label data, wherein the original label data comprises three types of labels of background, artery and vein; performing label conversion on the original label data, and converting three types of labels into two types of labels, namely background and blood vessel, to obtain converted image scanning data; preprocessing the voxel intensity of the converted image scanning data, cutting the voxel intensity value of the converted image scanning data to a preset range, and normalizing based on the mean value and variance of non-zero voxel intensities in scanning to obtain voxel processed scanning data; And spatially normalizing the scanned data after voxel processing to obtain normalized data, and resampling the normalized data according to a target distance to obtain labeled data and unlabeled data for semi-supervised training.
3. The pulmonary vessel CT image segmentation method based on semi-supervised learning as recited in claim 2, wherein the semi-supervised training framework includes a teacher model and a student model that are identical in structure and asynchronously updated in parameters, the training including a pre-training phase and a self-training phase; The training of the depth image segmentation model by adopting a multi-dimensional composite loss function based on a semi-supervised training framework to obtain a target image segmentation model specifically comprises the following steps: in a pre-training stage, training the student model by using labeled data, and initializing parameters of the teacher model through an index moving average mechanism; in a self-training stage, generating pseudo tag data with topological constraint through the teacher model for the non-tag data, and constructing the tag data and the pseudo tag data into mixed sample pairs through a feature mixing strategy; Training and updating parameters of the student model through a multi-dimensional composite loss function based on the mixed sample pair, and updating parameters of the teacher model through an index moving average mechanism based on the updated student model; the multi-dimensional composite loss function comprises a supervised loss calculated based on a mixed sample and a real label, an unsupervised loss calculated based on student model output and a pseudo label and a topological consistency loss calculated based on blood vessel skeleton similarity; And when the updated student model reaches a preset training stopping condition, taking the updated student model as a trained target image segmentation model and outputting the model.
4. A pulmonary vessel CT image segmentation method based on semi-supervised learning as recited in claim 3, wherein the supervised and unsupervised losses are both by a composite loss function And (3) calculating: ; Wherein, the Representing the cross-entropy loss, Representing the loss of the Dice similarity coefficient, Represents a loss of topological constraints based on the similarity of the vessel center skeleton, And a lambda_topo parameter in a weight coefficient representing the topology constraint loss.
5. The pulmonary vessel CT image segmentation method based on semi-supervised learning of claim 4, wherein the real label in the supervised loss computation is a real vessel label of labeling data, and the real label in the unsupervised loss computation is a pseudo label generated by the teacher model; the topological consistency loss is obtained by calculating the similarity between two vascular skeletons respectively predicted by the student model after the same label-free data are enhanced differently.
6. The pulmonary vessel CT image segmentation method based on semi-supervised learning according to claim 1, wherein the performing residual error superposition on the enhancement feature map of each stage and the up-sampling feature map of the corresponding stage of the decoder by the improved jump connection layer until the up-sampling of the decoder is completed, and outputting the segmentation result of the pulmonary vessel CT image, specifically includes: Inputting the bottom enhancement feature map to the decoder, and processing the bottom enhancement feature map through five cascade upsampling stages, wherein each upsampling stage sequentially comprises a transposed convolutional upsampling unit and a residual convolutional block; In a first upsampling stage, upsampling a bottom enhancement feature map by using the transposed convolution upsampling unit to obtain an upsampled feature map, performing residual superposition on the upsampled feature map and an enhancement feature map of a corresponding level encoder transmitted by the improved jump connection layer to obtain a fusion feature map, and processing the fusion feature map by using the residual convolution block to obtain an output feature map; In the remaining four upsampling stages, upsampling an output feature map of a previous upsampling stage through the transposed convolution upsampling unit to obtain a new upsampling feature map, performing residual superposition on the new upsampling feature map and an enhancement feature map of a corresponding level encoder transmitted through the improved jump connection layer to obtain a new fusion feature map, and processing the new fusion feature map through the residual convolution block to obtain an output feature map of a current upsampling stage; and taking the output characteristic diagram of the last upsampling stage as a final characteristic diagram, and adjusting the channel number of the final characteristic diagram through a 1X 1 convolution layer to obtain a final segmentation result.
7. A pulmonary vessel CT image segmentation system based on semi-supervised learning, wherein the pulmonary vessel CT image segmentation system based on semi-supervised learning is configured to implement the pulmonary vessel CT image segmentation method based on semi-supervised learning of any one of claims 1 to 6, the pulmonary vessel CT image segmentation system based on semi-supervised learning comprising: The model construction module is used for constructing a depth image segmentation model based on a U-shaped network architecture, and the depth image segmentation model comprises an improved encoder, an improved jump connection layer and a decoder; the semi-supervised training module is used for training the depth image segmentation model by adopting a multi-dimensional composite loss function based on a semi-supervised training framework to obtain a target image segmentation model; The model application module is used for acquiring a lung vessel CT image to be processed, inputting the lung vessel CT image into the improved encoder of the target image segmentation model for cascade downsampling and feature enhancement to obtain an enhancement feature map of each stage, and respectively carrying out residual superposition on the enhancement feature map of each stage and the upsampling feature map of the corresponding stage of the decoder through the improved jump connection layer until the upsampling of the decoder is completed, and outputting a segmentation result of the lung vessel CT image.
8. A terminal comprising a memory, a processor and a semi-supervised learning based pulmonary vessel CT image segmentation procedure stored on the memory and operable on the processor, the semi-supervised learning based pulmonary vessel CT image segmentation procedure, when executed by the processor, implementing the steps of the semi-supervised learning based pulmonary vessel CT image segmentation method of any of claims 1-6.
9. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a semi-supervised learning based pulmonary vessel CT image segmentation procedure, which when executed by a processor, implements the steps of the semi-supervised learning based pulmonary vessel CT image segmentation method according to any of claims 1-6.

Description

Lung vessel CT image segmentation method, system, terminal and storage medium based on semi-supervised learning Technical Field The invention relates to the technical field of image processing, in particular to a lung vessel CT image segmentation method, a system, a terminal and a computer readable storage medium based on semi-supervised learning. Background The blood vessel segmentation is a fundamental core problem in the field of medical image analysis, and accurate blood vessel segmentation can help doctors to quantify the stenosis degree, embolism position and range of pulmonary arteries, and a precise operation path is formulated for percutaneous saccule pulmonary angioplasty (BPA, percutaneous Pulmonary Balloon Angioplasty), so that the target blood vessel path searching time is shortened, the using amount of contrast agent and the radiation exposure risk of patients are reduced, and the operation safety and long-term prognosis effect are obviously improved. However, the prior art has a plurality of defects in vessel segmentation, such as uneven quality generated by pseudo labels, easy noise interference, insufficient capturing capability of a model on the characteristics of vessel subdivision branches, often the problems of fuzzy boundary judgment and branch distal end missed segmentation, poor cross-domain generalization capability and poor suitability on different vessel data sets. In addition, the existing partial enhancement module has the problems of insufficient suitability, such as high module calculation cost, limited gain aiming at a local tubular structure of a pulmonary blood vessel, easy loss of a thin blood vessel weak characteristic by a linear attention module, mismatching of a time sequence characteristic extraction module and a pulmonary blood vessel space structure segmentation requirement, and incapability of solving a core pain point. The existing loss function depends on Dice loss or cross entropy loss, only focuses on region overlapping degree, ignores vessel topological structure constraint, and leads to topological errors of segmentation results. Accordingly, the prior art is still in need of improvement and development. Disclosure of Invention The invention mainly aims to provide a pulmonary vessel CT image segmentation method, a pulmonary vessel CT image segmentation system, a pulmonary vessel CT image segmentation terminal and a pulmonary vessel CT image segmentation storage medium based on semi-supervised learning, and aims to solve the problems that noise interference is prone to occur in the existing pseudo labels in the prior art, boundary judgment is fuzzy during image processing, and the distal ends of branches are not segmented, so that the pulmonary vessel CT image segmentation precision is insufficient. In order to achieve the above purpose, the present invention provides a pulmonary vessel CT image segmentation method based on semi-supervised learning, which includes the following steps: Constructing a depth image segmentation model based on a U-shaped network architecture, wherein the depth image segmentation model comprises an improved encoder, an improved jump connection layer and a decoder; training the depth image segmentation model by adopting a multi-dimensional composite loss function based on a semi-supervised training framework to obtain a target image segmentation model; And acquiring a pulmonary vessel CT image to be processed, inputting the pulmonary vessel CT image to the improved encoder of the target image segmentation model for cascade downsampling and feature enhancement to obtain an enhancement feature image of each stage, and respectively carrying out residual superposition on the enhancement feature image of each stage and the upsampling feature image of the corresponding stage of the decoder through the improved jump connection layer until the upsampling of the decoder is completed, and outputting a segmentation result of the pulmonary vessel CT image. Optionally, in the pulmonary vessel CT image segmentation method based on semi-supervised learning, the training the depth image segmentation model by using a multi-dimensional composite loss function based on a semi-supervised training framework to obtain a target image segmentation model, the method further includes: Acquiring original medical image scanning data stored in an NIFTI format and corresponding original label data, wherein the original label data comprises three types of labels of background, artery and vein; performing label conversion on the original label data, and converting three types of labels into two types of labels, namely background and blood vessel, to obtain converted image scanning data; preprocessing the voxel intensity of the converted image scanning data, cutting the voxel intensity value of the converted image scanning data to a preset range, and normalizing based on the mean value and variance of non-zero voxel intensities in scanning to obtain voxel processed