Research On Domain Adaptation Methods For Medical Images

In medical image analysis,often we need to build an image recognition system for a target scenario,as well as multiple related models pretrained on different source scenarios.That refers to the process of transferring knowledge from multiple source domains to the target domain.However,this approach may result in a reduction in performance due to differences in data distributions across various scenarios.If labeled data is available in the target domain,this presents the combined challenges of multi-source-free domain adaptation and semi-supervised learning simultaneously.We present the problem for the first time and call it Semi-supervised Multi-source-free Domain Adaptation(SMDA).In order to effectively address both multi-source-free domain adaptation and semi-supervised learning in medical image analysis,novel algorithms and strategies need to be developed to make better use of limited labeled data and abundant unlabeled data,ultimately leading to improved image classification performance.However,both problems are typically studied independently in the literature,and how to effectively combine existing methods is non-trivial in design.In this work,we introduce a novel MetaTeacher framework with two key components:(1)Coordination weight learning for adaptive domain adaptation of individual source models.This means that the framework can adaptively transfer knowledge from multiple source domains to the target domain,which can overcome the limitation of the lack of labeled data in the target domain.(2)Bilevel optimization algorithm for consistently organizing the adaption of source models and the learning of target model.This enables the target model and the source models to interact with each other and adjust their learning process based on the feedback from each other.It aims to leverage the knowledge of source models adaptively whilst maximize their complementary benefits collectively.If there is no labeled data available in the target domain,it is referred to as MultiSource-free Domain Adaptation(MSDA).To tackle the challenge of the MSDA problem,this paper presents an extension of the MetaTeacher approach.A meta-learning based multi-source-free domain adaptation method is proposed.Specifically,we construct an adaptive confidence threshold within the original framework,a global confidence threshold as well as a local confidence threshold is constructed for each class,and dynamically adjust the threshold for each class during the meta-learning process.This generates pseudo-labeled samples that provide feedback signals for the teacher model instead of relying on labeled target domain data.The effectiveness of the methods was verified by constructing several transfer scenarios using five publicly available chest medical image datasets.