Entropy-Adversarial UDA with Contrastive Learning Boosts Brain Tumor Segmentation Across Centers

An unsupervised domain adaptation framework combining entropy-based adversarial alignment, pixel-wise contrastive learning, and channel-wise similarity regularization achieved 85.88% Dice on one private dataset and 86.62% on BraTS-2023 SSA for cross-center brain tumor segmentati…

• The EACC method integrates entropy-based global alignment, pixel-wise contrastive local discrimination, and channel-wise similarity to preserve structural consistency across domains.
• It outperformed all competing unsupervised domain adaptation techniques on private multi-center datasets, demonstrating practical value for real-world deployment.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.