Abstract
This study presents a novel Feature Pyramid Network with Dual-Decoder Supervision for accurate stroke lesion localization in multi-modal brain MRI. The proposed architecture integrates a Swin Transformer backbone with multi-scale feature aggregation, enabling effective fusion of hierarchical representations from DWI, ADC, and FLAIR sequences. A dual-decoder structure is employed, where the auxiliary decoder provides coarse lesion guidance through pseudo masks, and the primary decoder refines boundaries for precise voxel-level segmentation. Auxiliary supervision improves convergence stability and feature discrimination, while modality dropout enhances robustness to incomplete imaging protocols. Experiments conducted on the ATLAS v2.0 dataset demonstrate superior performance over baseline encoder–decoder models, achieving higher Dice scores, improved boundary accuracy, and strong lesion-wise detection rates. The model consistently localizes lesions of varying size, shape, and intensity, with minimal overfitting, as evidenced by small training–testing performance gaps. Qualitative results confirm the framework’s ability to transform coarse localization into anatomically accurate predictions. The combination of multi-modal integration, dual-decoder specialization, and self-training mechanisms positions the proposed method as a promising candidate for clinical deployment in rapid stroke diagnosis workflows. Future directions include expanding validation to multi-center datasets, incorporating explainable AI techniques, and enabling real-time 3D processing for deployment in acute care environments.