| With the development of medical image visualization technology,medical images is playing an increasingly important role in clinical assistant diagnosis analysis.However,the large number of MRI and CT images that need to be reviewed places high demands on doctors’ diagnostic efficiency,which is well alleviated by medical image processing and analysis based on computer vision,image processing and analysis techniques,and artificial intelligence technologies.And medical image segmentation is an important part of medical image processing and analysis.Therefore,the research on highly robust and high-precision medical image segmentation methods has extensive and important clinical application value.Aiming at difficult problems in medical image multi-objective joint segmentation task,such as gray similarity of different organs,different gray representation of the same organ,noise and so on,this dissertation analyzes the characteristics and advantages of traditional machine learning methods and deep learning methods,investigates the characteristics of advanced technologies,designs medical image multi-objective joint segmentation methods to solve different difficult problems,and carries out validation and analysis on public datasets Synapse and ACDC.Experimental verification and analysis are carried out on the Synapse and ACDC datasets.The specific research content and innovation work are summarized as follows:(1)Aiming at the difficult segmentation problems of gray similarity of different organs,different gray representation of the same organ and noise in 3D medical images,a 3D medical image segmentation method based on joint deep network and morphological structure constraints is proposed.This method is based on Transformer and CNN combined with morphological structure constraint information,and leverages shape prior information of tissues and organs to constrain the segmentation results,enhancing the robustness and interpretability of the multi-objective joint segmentation model.The results of comparison experiments and ablation studies show that the method can improve the over-segmentation and under-segmentation problems of medical image segmentation,and enhance the performance of multi-objective joint segmentation.(2)In order to solve the limitation of the utilization of image contextual information and underlying semantic information caused by the skip-connection in U-shaped network,a reticular Transformer(RT)with the collaboration of information aggregation for 3D medical image segmentation method is proposed.Based on Swin-Unet,this method builds a novel structure for aggregation of multi-scale features,offsetting the drawbacks of skip-connection by extracting texture information and underlying semantic information.Meanwhile,the attention mechanism is utilized to realize the better weight allocation of features.The results of comparison experiments and ablation studies show that this method can improve the accuracy of multi-objective joint segmentation,accelerate the convergence rate of the model and enhance the interpretability of the model.(3)In order to further optimize the multi-objective joint segmentation model for the difficult problem of segmentation boundary ambiguity,a 3D medical image segmentation method based on the causal inference decoupling Transformer is proposed.This method applies the theory of causal inference to the image segmentation task,constructs causal relationships between tissues and organs from the perspective of causal analysis and uses them as a kind of a prior information,removes features and false associations that are irrelevant to segmentation targets,and focuses on the essential features of segmentation targets to make the segmentation of targets more accurate.Experimental results verify this method can better detect edges of targets,and enhance the edge detection performance of multi-objective joint segmentation,as well as solve the task of multi-objective joint segmentation of medical images more effectively. |
