Cross-Center Domain Adaptive Peripheral Edema Identification Based on Adversarial Learning and Regional Attention

Addressing the task of accurate identification of peripheral edema in brain diseases, existing methods mostly rely on unimodal or whole-brain feature extraction, failing to take full advantage of the complementarity of multimodal data and key features of the microenvironment surrounding the lesion. Meanwhile, distributional differences in cross-institutional data further limit the generalisation ability of the model. Therefore, a cross-center domain adaptive framework, namely ADRA-Net, based on adversarial learning and regional attention, is proposed to address the cross-domain distributional differences and improve the identification performance of edema regions. In this study, a total of 274 subjects from multiple scanners across institutions are collected to identify edema regions in brain metastases (BM) and central nervous system infections (CNSI). The experimental results show that ADRA-Net achieves significant classification performance on DWI modality sequences for cross-center brain disease peripheral edema, which can reach 83.7% and 81.3% on the validation set and external test set. We further validate the effectiveness of the components through ablation experiments, providing an effective solution for medical image analysis in complex scenarios.