TestFit: A plug-and-play one-pass test time method for medical image segmentation

Yizhe Zhang, Tao Zhou, Yuhui Tao, Shuo Wang, Ye Wu, Benyuan Liu, Pengfei Gu, Qiang Chen, Danny Z. Chen

Abstract: Highlights•TestFit is a plug-and-play one-pass test-time training method for off-the-shelf segmentation networks.•The model update in TestFit is performed in an online fine-tuning fashion, which requires no extra training data, no new network design, and no offline retraining.•TestFit effectively prevents model collapse issues and achieves a good balance between the stability and effectiveness of test-time training.