Open-Set Source-Free Domain Adaptation in Fundus Images Analysis
Abstract
Unsupervised domain adaptation (UDA) is crucial in medical image analysis where only the source domain data is labeled. There is a lot of emphasis on the closed-set paradigm in UDA, where the label space is assumed to be the same in all domains. However, medical imaging often has an open-world scenario where the source domain has a limited number of disease categories and the target domain has unknown distinct classes. Also, maintaining the privacy of patients is a crucial aspect of medical research and practice. In this work, we shed light on the Open-Set Domain Adaptation (OSDA) setting on fundus image analysis while preserving the privacy concern. In particular, we step towards a source-free open-set domain adaptation where, without source data, the source model is utilized to facilitate adaptation to open-set unlabeled data by delving into channel-wise and local features for fundus disease recognition. In particular, considering the nature of the fundus images, we present a novel objective way in the adaptation phase to utilize spatial and channel-wise information to select the best source model for a target domain, even by considering the small inter-class variation between samples. Our approach has achieved state-of-the-art performance compared to other methods.