Research On Representation Optimized Self-supervised Medical Image Segmentation Based On Siamese Network

The segmentation of organs,tissues and other targets in medical images plays an important role in medical diagnosis,treatment and prognosis.Building an accurate and robust automatic medical image segmentation system has far-reaching research significance.However,the establishment of a large-scale annotated medical dataset requires high-level medical expert knowledge,and dense point annotation is costly on time.On the unannotated medical image data,the application of self-supervised learning to improve the segmentation performance has been in-depth research and development.However,most previous studies designed proxy task training and seldom paid attention to the representation learning of the medical image itself in the latent space.Moreover,the scale diversity of medical segmentation targets is ignored.Aiming at this problem,the dissertation effectively combines the existing global,local and dense representations in self-supervised Siamese network representation learning and implements them in the field of medical images,and proposes a self-supervised medical image segmentation method with multi-scale representation consistency.Specifically,the dissertation mainly conducted the following research:Firstly,the generation of differential views required for the self-supervision of the medical image Siamese network paradigm is investigated.Aiming at the problems of limited application scope and high computational complexity of existing alignment methods,a canvas matching method is proposed,which can obtain richer and more complex spatial transformation while the dense location correspondence information is efficiently obtained.Secondly,the existing Siamese network representation extraction is researched,and the multi-scale representation extraction Siamese network is designed,in which the multiscale projection representation and prediction representation are obtained through the convolutional form projectors and predictors corresponding to each scale.Among them,in order to solve the problem of excessively large hidden dimension tensor caused by dense scale representation operations,an embedding pre-sampling module is proposed to ensure efficient training.Finally,the learning loss of multi-scale representations is studied.Aiming at the problems of individual differences and scaling differences in medical image target scales,a representation consistency loss is proposed to automatically match the best scales between representations to learn scale invariance.At the same time,in order to solve the problem of the lack of contextual information on the view edge,the centrality weight is proposed to reduce the influence of representations on the view edge.The dissertation implements multi-scale representation learning on the RibFrac dataset and multiple three-dimensional CT datasets on Medical Segmentation Decathlon,and performs downstream segmentation experiment on the challenging multi-organ segmentation Beyond the Cranial Vault dataset of Vanderbilt University Medical Center for performance verification.The experiment is compared with the existing advanced medical segmentation network and a variety of different scale representation learning methods,as well as improvement independence analysis and annotation efforts analysis to verify the effectiveness and robustness of the self-supervised medical image segmentation method with multi-scale representation consistency proposed in the dissertation.