Research On Medical Image Segmentation Algorithm Based On Deep Learning Under Incomplete Supervision

Medical image segmentation aims to delineate different organs or lesion,which plays a crucial role in clinical diagnosis,quantitative analysis of tissue volumes,depiction of anatomical structures and surgical planning.Relying on sufficient annotated data,deep learning methods have achieved significant progress in medical image segmentation tasks.However,it is time-consuming and laborious to draw the outline of target regions in medical images in clinical practice.The labeling process requires domain knowledge,which makes it expensive and difficult to obtain sufficient and well-labeled datasets.Therefore,most of the clinical situations are under an incomplete supervised condition,i.e.,there exists a small amount of annotated data and a large amount of unannotated data to use.Under the incomplete supervision condition,the expansion of labeled data and the mining of unlabeled data are crucial solutions to improve the segmentation performance.In this thesis,we focus on how to design methods for efficient expansion of labeled data and effective mining of unlabeled data,which aims to improve the segmentation performance.Further,we verify the superiority of the proposed methods under the incomplete supervision condition through extensive experiments.The specific contributions of this thesis are summarized as the following four aspects:To address the labeling sample selection problem in segmentation tasks under the incomplete supervision,we propose a quality assessment driven active learning method partialsupervised segmentation in this thesis.Since the criteria of uncertainty and representativeness,which are commonly used in active learning,usually cannot accurately reflect the segmentation performance of the unlabeled samples.Thus,it is difficult to select the unlabeled samples with poor segmentation quality.We evaluate the segmentation quality based on the attention mechanism and deep supervised loss.The quality assessment driven active learning guides the sparse annotation process.Then,the segmentation network is trained under partial supervision.The proposed quality assessment driven active learning method only requires a small amount of annotation queries to achieve comparable results to fully-supervised training and outperform other annotation strategies.To mitigate the adverse effect brought by the noisy components in pseudo labels under incomplete supervision.This thesis proposes an error correction guided semi-supervised segmentation method.This thesis is the first attempt to divide the segmentation errors into intra-class error and inter-class error.Further,the specific mask degradation methods are designed to highlight the typicality of the segmentation errors.Then,a discriminative segmentation error prediction network by collaborative multi-task learning is proposed to predict segmentation errors.Further,a two-stage error rectification method is designed to guide the semi-supervised segmentation self-training.Extensive experiments conducted on two colon gland segmentation datasets demonstrate that the segmentation error predictions can improve the initial performance of different baseline networks.The error correction guided semisupervised self-training achieves better performance than other popular semi-supervised segmentation methods.To address the unstability and poor generalization performance caused by insufficient feature discrimination in semi-supervised self-training.This thesis proposes a dynamic prototypical feature representation learning method for the self-training of semi-supervised segmentation.A color-exchange based pseudo-label generation method is designed to provide reliable supervised information for segmentation feature representation learning.Then,a memory relational learning method is developed to boost the feature representation globally.Further,a confidence-aware contrast learning method is introduced.The prototypical classifier is used to ensure the reliability of the pair construction in semi-supervised contrast learning.The pixel-wise feature representation is enhanced locally.Extensive experiments on four medical image segmentation datasets demonstrate that the proposed method achieves better semi-supervised segmentation performance,outperforming other popular semi-supervised segmentation methods in recent years.The segmentation results of the proposed method is close to the performance achieved by fully-supervised training.To address the poor training stability and low learning efficiency caused by the mismatched feature distributions between unlabeled data and labeled data and insufficient feature discrimination in semi-supervised consistency learning.This thesis proposes an implicit and explicit prototype alignment consistency learning method for semi-supervised segmentation.The implicit and explicit prototype alignment enhance the distribution matching between labeled data and unlabeled data,and enhance the feature discrimination simultaneously.In implicit alignment,the multiple segmentation masks are generated by the prototypical feature matching with multiple online-generated prototypes.The diversity and stability in consistency learning are enhanced in this way.Further,a voting strategy based on multiple predictions is introduced to enhance the reliability of unlabeled masks for prototype calculation.In explicit prototype alignment,we construct three hierarchical explicit prototype alignment tasks from labeled to unlabeled and from certain to uncertain regions,enhancing intra-class tightness and inter-class separability of segmentation feature representations.Extensive experiments on four medical image segmentation datasets validate the effectiveness of the proposed method.