EACC: An Entropy-Based Adversarial UDA Network with Inter-domain Channel-Wise Similarity Regularization and Pixel-Wise Contrastive Learning.

Magnetic resonance images acquired from different medical centers often exhibit substantial variations in imaging protocols and scanners, resulting in severe domain shifts that degrade the performance of brain tumor segmentation models. This study develops an effective unsupervised domain adaptation (UDA) framework for cross-center brain tumor segmentation, enabling models trained on labeled source data to generalize to unlabeled target datasets collected from different institutions. We propose an entropy-based adversarial UDA network with inter-domain channel-wise similarity regularization and pixel-wise contrastive learning (EACC). Specifically, entropy-based adversarial learning performs global domain alignment, pixel-wise contrastive learning enhances local feature discrimination, and channel-wise similarity regularization preserves structural consistency across domains. These components collaboratively reduce domain discrepancies, thereby improving cross-center generalization. Experiments were conducted on two private datasets, Brain-FZ129 and Brain-ZZ211, using BraTS2020 as the source domain. On Brain-FZ129, EACC achieved 81.82% in Dice, 6.45 mm in 95% Hausdorff Distance (HD95), 69.38% in Jaccard Index, and 2.18 mm in average surface distance, outperforming all competing UDA methods. On Brain-ZZ211, EACC also obtained the best performance, with a Dice of 85.88% and an HD95 of 6.56 mm, ranking first among all compared methods. Furthermore, on the public BraTS-2023 SSA dataset, our method achieved 86.62% in Dice and 15.84 mm in HD95, demonstrating competitive performance and confirming the generalization capability of the proposed approach. These results indicate that EACC effectively reduces domain shifts and improves the robustness of brain tumor segmentation across heterogeneous medical centers, providing practical value for real-word clinical deployment.