Continual source-free active domain adaptation for nasopharyngeal carcinoma tumor segmentation across multiple hospitals.

Abstract

The Nasopharyngeal carcinoma (NPC) is a common malignant tumor and precise gross tumor volume (GTV) delineation is crucial for effective NPC radiotherapy. Deep learning techniques have enabled automated GTV segmentation, nevertheless, model performance often degrades due to domain shifts in multi-center data scenarios. Recent source-free active domain adaptation methods have achieved promising results; however, these are still limited by several issues: (i) dependency on the source data features, (ii) inappropriate selection of biased and redundant samples, and (iii) catastrophic forgetting. In the current investigation, we propose a novel continual source-free active domain adaptation (CSFADA) framework for GTV segmentation of NPC. Inspired by self-supervised and cross-correlation learning, we introduce a domain reference and invariants selection strategy to address the first two challenges mentioned above. To this end, the strategy first acquire target domain knowledge in a self-supervised learning manner. It then computes a domain-distance and a domain-invariance score for each sample, thereby selecting informative samples. To address the third challenge mentioned above, we develop a dual-stage recurrent distillation strategy based on the clinical practice. Specifically, the stage I employs self-supervised learning approaches to learn generalizable representations and preserve source domain knowledge. The stage II decouples classical knowledge distillation to avoid optimization conflicts and thus better preserve source domain information. We conduct extensive experiments on datasets from three centers for GTV segmentation of NPC. The experimental results demonstrate the superiority of our proposed methods. Our code is publicly available at https://github.com/YZC-99/CSFADA.git.