Genetic Algorithm-Enhanced Fusion Network for Multi-Sequence MRI Brain Tumor Characterization

Accurate brain tumor classification from MRI scans is crucial for treatment decisions and patient outcomes. This paper introduces GA-EFN, a network that combines information from different MRI sequences (T1-weighted, T1-contrast, and T2-weighted images) to improve tumor classification. The system uses three ResNet-34 networks to process each MRI type separately, then merges them using an adaptive fusion module optimized through genetic algorithms. This approach automatically finds the best way to combine features that distinguish between tumor types. Testing on a dataset with 17 tumor classes showed impressive results: 97.4% accuracy and 96.8% macro-F1 score with particularly strong performance on gliomas, pituitary tumors, and metastatic cases. The genetic algorithm optimization proved effective, consistently improving validation scores over successive generations. Notably, the fusion approach remained robust even when individual MRI sequences were corrupted or missing, outperforming both singlesequence methods and standard fusion techniques. These findings suggest that combining evolutionary optimization with multi-sequence MRI analysis offers a promising path for automated brain tumor diagnosis in clinical settings.