FedCAF: Federated Confidence-Adaptive Fusion Framework for Dual-Modal CT and MRI Brain Tumor Classification

Abstract

Automated brain-tumor classification from medical pictures is complex issue because CT and MRI provide various and complimentary representations of brain abnormalities, and also medical imaging data are widely spread across institutions due to privacy and data-sharing constraints. Most of the existing works were on single modality categorization, traditional multi-modal fusion or federated learning separately. Here, we propose a dual-modal CT and MRI brain-tumor classification via a federated confidence-adaptive system. The proposed FedCAF framework adopts a modality-aware dual-backbone architecture using EfficientNetV2-S for CT image feature extraction and Swin-Small for MRI image feature extraction. The extracted representations are projected onto a shared feature space and processed by employing Dynamic Hierarchical Confidence Estimation (D-HCES) to assess the reliability of features. The generated confidence information is used to direct the Confidence-Adaptive Fusion Module (CAFM) to adaptively weight the features. Then the revised representation is processed by the Dual-Gated Confidence-Guided Self-Attention Module (DGCGSAM), and is classified in binary classification. The full model is trained utilizing Federated Averaging (FedAvg) on several distributed clients, where the model parameters are transmitted between clients and the central server, rather than raw medical images. The suggested FedCAF framework achieved an overall accuracy of 96.74%, with a precision of  99.87%,recall of 94.24%,and F1-score of  96.97%  on the final test set. The federated training over four clients showed a successful collaborative optimization while the medical images remained local. The results show that the proposed framework helps in reliable classification of brain tumors using CT and MRI images.