Cross-Dataset Deep Learning for Robust Brain Disease Detection and Classification Using Multimodal MRI

Abstract

Brain disease detection and classification using magnetic resonance imaging (MRI) remain challenging because variations in scanners, acquisition protocols, imaging modalities, and disease-specific characteristics introduce domain shifts that reduce deep learning generalizability. This study develops a robust cross-dataset framework for automated brain disease detection and multiclass classification using heterogeneous MRI data. Public benchmark datasets, including BraTS, ADNI, and OASIS, are utilized to represent different brain pathological conditions, with dataset-specific MRI sequences subjected to harmonized preprocessing and subject-level partitioning. The proposed Cross-Dataset Multimodal Attention Fusion Network (CD-MAFNet) integrates modality-specific convolutional encoders, attention-guided feature fusion, and domain-invariant representation learning to extract disease-relevant features while reducing dataset-specific biases. MRI scans are normalized, spatially aligned, resized to 224 × 224 pixels, and augmented. Training employs Adam with a 0.0001 learning rate, batch size of 32, and 100 epochs. Within-dataset evaluation achieves 96.84% accuracy, while unseen cross-dataset testing obtains 92.46% accuracy, 91.83% F1-score, and 95.37% AUC, with a 4.38-percentage-point generalization gap. The proposed approach improves cross-dataset accuracy by 3.21 percentage points over the strongest baseline. The novelty lies in integrating multimodal attention fusion with domain-invariant learning for cross-dataset brain disease classification. The findings demonstrate improved robustness against inter-dataset variations and support reliable, generalizable, computer-assisted brain disease diagnosis from MRI.