Research And Implementation Of Label Efficient Medical Image Segmentation Algorithm Based On Prototype Learning

3D CT medical image segmentation is the process of extracting the target 3D structure from the region of interest(Region of Interest,ROI)image corresponding to the region where the organ or tissue is located.In the past few years,many work have proposed fully automatic and semi-automatic image segmentation methods based on deep learning,but they often require a large amount of labeled data to support learning,and are difficult to achieve good results in medical image segmentation tasks that lack fine annotations.Therefore,this thesis proposes a novel semi-supervised learning method based on prototype learning,in which the recurrent prototype learning paradigm enables the model to learn the correct segmentation algorithm directly from the few-shot dataset without pre-training,which improves label efficiency of the model and provides the possibility for the clinical application of deep learning-based algorithms.The specific work of this thesis is as follows:(1)This thesis designs a cycle prototype learning model(Cycle Prototype Learning,CPL)for medical image segmentation,which includes two prototype learning processes,forward process and reverse process.The forward process uses the labeled support image extraction prototype to segment the unlabeled query image,and generates pseudo-labels of the query image through the Prior Pseudo Optimization Module(PPOM)to learn information from it;the reverse process uses Using the prediction of the query image in the forward process as the label to extract the prototype,try to recover the supporting image label.The experimental results based on the kidney and brain CT medical image datasets show that,compared with the classic end-to-end fully supervised segmentation network,the CPL designed in this chapter achieves better results in the context of few-shot medical image segmentation.(2)This thesis designs a fine decoding network(Fine Decoding Feature Extractor,FDFE)for fine-grained feature extraction,and embeds it in CPL to form a few-shot semi-supervised medical image segmentation model Cycle Prototype Network(CPNet),which further improves the segmentation accuracy.FDFE supplements detailed features lost during resolution compression with cross-layer feature transfer,and guarantees distortion-free feature embeddings through coordinate inheritance up-pooling.Experiments on kidney and brain datasets show that CPNet has significantly improved performance compared to the CPL proposed above.Detailed ablation studies further demonstrate the contribution of each module in CPNet to algorithm performance.(3)On the basis of the above medical image segmentation algorithm,this thesis implements the corresponding segmentation system,so that the algorithm can be put into use in the future to meet the needs of automatic segmentation of medical images.The system uses the Pythonbased Py Qt5 package to implement the function of multi-functional medical image segmentation,and writes an interface-based image browsing and dumping function.