Decoder-Free Supervoxel GNN for Accurate Brain-Tumor Localization in Multi-Modal MRI 

Most MRI-based tumor models rely on encoder–decoder pipelines, where a large portion of parameters is devoted to voxel-wise reconstruction. We took a different approach: remove the decoder entirely and concentrate learning capacity where it matters most, in the encoder. Our model, SVGFormer, turns multi-modal MRI into a supervoxel graph. Local image patches feed a Transformer-based embedder, while a GNN captures context across regions. Specialized heads then localize tumors directly at the node level. With no decoder, every parameter is dedicated to richer feature learning and region-level reasoning.

Our role in this project

We achieved strong tumor localization on BraTS 2025, with an F1 score of about 0.88, a ROC-AUC of about 0.98, and a MAE of about 0.03. Our approach provides dual-scale interpretability, ranging from graph-level attention down to individual voxel patches. Moreover, the decoder-free design channels model capacity into the encoder, which is the proven driver of performance.