Improving Brain Tumor Segmentation with a Clinically-Informed Multi-Decoder U-Net

Motivation: Advancements in deep learning have improved brain tumor segmentation, but many methods do not leverage clinically relevant input modalities for specific tumor subtypes. Goal(s): We aimed to develop a multi-decoder U-Net architecture to enhance segmentation accuracy by integrating modality-specific MRI sequences. Approach: Our approach utilized a shared encoder branching into three decoders, each tailored to a tumor region, incorporating guided squeeze-and-excitation attention blocks for optimal feature integration. Results: Evaluated on multiple datasets, our model showed significant improvements in Dice scores —up to 2.5% on the BraTS Africa dataset —demonstrating robustness and generalizability. Impact: Our clinically guided, multi-decoder U-Net demonstrates improved segmentation accuracy, particularly in diverse and non-standard datasets. This innovation paves the way for more reliable, adaptable, and interpretable brain tumor imaging, enhancing diagnostic confidence and treatment planning.