Reproduce “From Denoising Training to Test-Time Adaptation: Enhancing Domain Generalization for Medical Image Segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruxue Wen, Hangjie Yuan, Dong Ni, Wenbo Xiao, Yaoyao Wu
Researchers at Zhejiang University developed DeY-Net, a U-Net based architecture with a self-supervised denoising branch, coupled with Denoising Test-Time Adaptation (DeTTA) to enhance medical image segmentation on unseen domains. The approach achieved a 1.99% average Dice coefficient improvement over baseline U-Net on various CT datasets, demonstrating robust generalization from a single source domain.
Key insights
- Self-supervised denoising, particularly through N2V, effectively guides the network to learn more robust and domain-invariant features by focusing on underlying clean image structures, even from noisy medical data.
- Integrating an auxiliary denoising task within a Y-shaped network architecture allows for joint optimization of segmentation and denoising, improving generalization without introducing excessive complexity.
- Test-Time Adaptation, specifically adapting Batch Normalization layers with a denoising objective and leveraging ensemble predictions from noise-corrupted inputs, offers a powerful mechanism to dynamically adjust models to unseen target distributions at inference time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.