Skip to main navigation Skip to search Skip to main content

Day–Night Adaptation: A domain adaptation framework for medical image segmentation without source data

  • Ziyang Chen
  • , Yiwen Ye
  • , Yongsheng Pan
  • , Jingfeng Zhang
  • , Yong Xia
  • Northwestern Polytechnical University Xian
  • Ningbo No. 2 Hospital

Research output: Contribution to journal › Article › peer-review

Abstract

Medical image segmentation models often suffer from performance degradation when deployed across clinical centers due to distribution shifts arising from differences in scanners, acquisition protocols, and patient populations. This challenge limits the robustness of deep models in real-world healthcare, where strong cross-center generalization is critical. Conventional unsupervised domain adaptation alleviates such shifts by jointly accessing source and target data, but this strategy is incompatible with strict privacy regulations that prohibit inter-center data sharing. More recent paradigms, including source-free domain adaptation (SFDA) and test-time adaptation (TTA), attempt to preserve privacy by relying only on target data. However, SFDA assumes a pre-collected target dataset prior to deployment, which is often impractical in clinical workflows, while TTA processes each case only once, underutilizing the accumulating data and risking unstable adaptation. To address these limitations, we propose Day-Night Adaptation (DyNA), a clinically motivated, continuous adaptation framework that operates in synchrony with routine hospital schedules. During the day, DyNA performs adaptation on each test sample by learning a low-frequency prompt on the Fourier spectrum while keeping the backbone frozen, aided by a memory bank for initialization and a warm-up mechanism for stable optimization. At night, DyNA reuses the stored test samples collected during the day to conduct self-training with an improved teacher–student paradigm, where a global student bridges the gap between the EMA-updated teacher and the rapidly adapting student, enabling stable fine-tuning. In this way, DyNA unifies the advantages of online and offline adaptation, achieving reliable performance improvement without violating privacy constraints or requiring pre-collected target datasets. Extensive experiments on two benchmark medical image segmentation tasks demonstrate that DyNA consistently outperforms state-of-the-art SFDA and TTA methods. These results highlight the potential of DyNA as a privacy-preserving, deployment-oriented strategy for continual cross-center adaptation in real clinical environments. Codes and pre-trained source models will be publicly released at https://github.com/Chen-Ziyang/DyNA.

Original languageEnglish
Article number114576
JournalPattern Recognition
Volume180
DOIs
StatePublished - Dec 2026

Keywords

  • Distribution shift
  • Domain adaptation
  • Medical image segmentation

Fingerprint

Dive into the research topics of 'Day–Night Adaptation: A domain adaptation framework for medical image segmentation without source data'. Together they form a unique fingerprint.

Cite this