Robust Source-Free Domain Adaptation for Medical Image Segmentation based on Curriculum Learning

TL;DR

This work addresses the privacy-sensitive setting of source-free unsupervised domain adaptation for medical image segmentation. It introduces Learning from Curriculum (LFC), which combines an easy-to-hard curriculum (guided by prediction-consistency measured via $d = KL(p^s,p^t)$) with a source-to-target curriculum built on a triplet-branch model and a momentum network, aided by a self-supervised augmentation-consistency loss to exploit unlabeled target data. The learning process blends fixed source pseudo-labels with progressively updated pseudo-labels from target data through a smooth update, improving stability during adaptation. Experiments on cross-domain fundus (OD/OC) and polyp segmentation demonstrate state-of-the-art performance among source-free methods, highlighting the practical impact for privacy-preserving medical image analysis.

Abstract

Recent studies have uncovered a new research line, namely source-free domain adaptation, which adapts a model to target domains without using the source data. Such a setting can address the concerns on data privacy and security issues of medical images. However, current source-free domain adaptation frameworks mainly focus on the pseudo label refinement for target data without the consideration of learning procedure. Indeed, a progressive learning process from source to target domain will benefit the knowledge transfer during model adaptation. To this end, we propose a curriculum-based framework, namely learning from curriculum (LFC), for source-free domain adaptation, which consists of easy-to-hard and source-to-target curricula. Concretely, the former curriculum enables the framework to start learning with `easy' samples and gradually tune the optimization direction of model adaption by increasing the sample difficulty. While, the latter can stablize the adaptation process, which ensures smooth transfer of the model from the source domain to the target. We evaluate the proposed source-free domain adaptation approach on the public cross-domain datasets for fundus segmentation and polyp segmentation. The extensive experimental results show that our framework surpasses the existing approaches and achieves a new state-of-the-art.

Figures (2)

• Figure 1: Pipeline of our learning from curriculum (LFC) framework. The proposed LFC consists of two curricula, i.e., easy-to-hard (top) and source-to-target (bottom).
• Figure 2: Exemplars from fundus and colonoscopic image datasets adopted in our study. The source domains (i.e., REFUGE and CVC-Clinic) are visually different from the target ones (known as domain shift).