The impact of U-Net architecture choices and skip connections on the robustness of segmentation across texture variations

This work was published in Computers in Biology and Medicine (CiBM), volume 197, article 111056, 2025.

Read the paper: Publisher (Computers in Biology and Medicine) · Google Scholar

The U-Net has become the default architecture for medical image segmentation, and more than a hundred variants have been proposed — mostly modifying its skip connections, backbones, and bottlenecks. Skip connections are widely believed to carry fine spatial detail across the encoder-decoder gap, but how much they help, and how they interact with robustness under real-world variability, is not well understood.

This journal paper extends the MICCAI 2023 conference study (see Chapter 10) with a broader analysis across task complexities and texture variations. We find that the benefit of skip connections is small for low-to-medium complexity tasks and grows only as task complexity becomes large, and — critically — that advanced variants such as the Attention-Gated U-Net and UNETR are not consistently more robust than a standard U-Net under distribution shifts. NoSkip architectures, or those using addition-based skip connections (e.g., V-Net), offer better stability on out-of-domain data, while dataset diversity (texture variation and foreground-background balance) further improves generalizability. These results argue for careful, robustness-aware design of skip connections rather than adopting more complex architectures by default.

Citation

If you find this work useful, please cite it as:

@article{kamath2025impact,
title={The impact of U-Net architecture choices and skip connections on the robustness of segmentation across texture variations},
author={Kamath, Amith and Willmann, Jonas and Andratschke, Nicolaus and Reyes, Mauricio},
journal={Computers in Biology and Medicine},
volume={197},
pages={111056},
publisher={Elsevier},
doi={10.1016/j.compbiomed.2025.111056},
year={2025}
}