Research On Medical Image Segmentation Based On Domain Adaptation

Medical image segmentation is an important part of smart medical care that involves the automatic identification of target pixels,such as organs,tissues,or lesions in med-ical images,to aid doctors in the diagnosis and treatment of diseases.This technology has broad prospects in clinical diagnosis,auxiliary therapy,and medical research.Do-main adaptation are commonly used to address distribution shift problem among medical image datasets.However,the existing medical image segmentation algorithms based on domain adaptation still have the following problems in practical applications:(1)The neg-ative transfer problem in knowledge transfer?(2)The error diffusion problem in source-free cross-domain scenarios?(3)The homogeneous feature extraction problem in cross-modality scenarios.These problems has been addressed via following efforts:1.For the basic cross-domain medical image segmentation scenarios,aiming at the problem that existing algorithms will introduce negative transfer,resulting in a degra-dation of model segmentation performance,this thesis proposes an unsupervised domain adaptation medical image segmentation algorithm based on adversarial-joint training(Ad-versarial Joint-Training Domain Adaptation for Medical Image Segmentation,AJTDA).AJTDA uses the pre-training strategy to enhance the convergence ability of the model,and avoids negative transfer through adversarial joint training under the premise of retaining the prior distribution information of the source medical images.Compared with existing algorithms,the experimental results of cross-domain organ segmentation on five public datasets show that the segmentation performance of the AJTDA has improved by 1.72%,3.35%,2.36%,0.59%,2.54%,respectively.2.For the source-free cross-domain medical image segmentation scenarios,aim-ing at the problem that existing algorithms will lead to error diffusion,this thesis pro-poses a source-free cross-domain medical image segmentation algorithm based on reduc-ing model style sensitivity(Style-Insensitive Source-Free Domain Adaptation for Medi-cal Image Segmentation,SI-SFDA).SI-SFDA improves the generalization ability of the source model via the style generalization mechanism,and further addresses the error dif-fusion problem caused by the style shifts via the adaptive confidence regularization loss.Compared with existing algorithms,the experimental results of source-free cross-domain organ segmentation on five public datasets show that the segmentation performance of the SI-SFDA has improved by 2.83%,2.64%,3.21%,3.01%,3.32%,respectively.3.For the cross-modality medical image segmentation scenarios,aiming at the prob-lem that existing algorithms cannot extract homogeneous features between modalities,resulting in difficulty to align the data distributions,this thesis proposes a cross-modality medical image segmentation algorithm based on soft sharing synergistic adaptation(Soft-Shared Synergistic Domain Adaptation for Cross-Modality Medical Image Segmentation,S~3CMDA).S~3CMDA extracts the homogeneous features of medical images between dif-ferent modalities by the soft sharing mechanism,and on this basis,solves the distribution shift problem by a synergistic adaptation strategy.Compared with the existing algorithms,the results of two sets of cross-modality segmentation experiments on the public heart segmentation dataset Cardiac show that the segmentation performance of S~3CMDA has improved by 0.92%and 0.75%,respectively.