RIOG: Rectify-to-Match Gradient for Source-Free Domain Adaptive Medical Image Segmentation

Source-free domain adaptation (SFDA) facilitates knowledge transfer from source models to unlabeled target data with inaccessible source data. An auxiliary task-based self-training paradigm is typically employed to update source models by selecting appropriate auxiliary tasks (e.g., consistency constraints). However, these approaches face several problems: 1) objective discrepancies between auxiliary tasks and segmentation tasks can cause gradient deviation of source models, leading to dramatic performance degradation and 2) significant class imbalance in medical images often causes gradient optimization biased toward the majority classes, resulting in misclassification of minority classes. In this article, we propose a novel rectify-to-match gradient (RIOG) framework for SFDA medical image segmentation. First, we develop the gradient rectification network to effectively learn the gradient deviation, thereby bridging the gap between auxiliary tasks and desirable segmentation tasks during the target-domain adaptation. Second, we introduce a class-balance optimization strategy that matches the gradient magnitude and direction between majority and minority classes to mitigate class imbalance. Extensive experiments on fundus image and brain tumor segmentation tasks demonstrate the effectiveness of the proposed RIOG, establishing its superiority over state-of-the-art SFDA methods. The code will be available at https://github.com/dogeONE-bit/RIOG