A Supervoxel Graph-Transformer Framework for Lightweight and Interpretable Brain Tumor Segmentation from Multimodal Magnetic Resonance Imaging.
Authors
Affiliations (4)
Affiliations (4)
- African Institute for Mathematical Science (AIMS), Mbour, Mbour, 1418, Senegal.
- Computer Engineering, University of Yaounde I, Yaounde, Yaoundé, Centre Region, 8390, Cameroon.
- AIMS Senegal, Mbour, Mbour, Thiès Region, 1418, Senegal.
- Saint Petersburg State University, Saint pertersburg, saint, 7/9, Russian Federation.
Abstract
Accurate segmentation of brain tumor sub-regions from multimodal MRI is essential for diagnosis and treatment planning, but most deep learning methods process millions of voxels, which is computationally demanding. We propose TransformerGNN, a supervoxel graph-Transformer framework that reformulates volumetric segmentation as node classification on a sparse graph. Each multimodal volume (FLAIR, T1, T1ce, T2) is partitioned into approximately 2,000 SLIC supervoxels, reducing about 8.9 million voxels to a region adjacency graph whose nodes carry multimodal intensity and spatial features. Stacked TransformerConv layers combine graph connectivity with multi-head attention. On BraTS 2020, the model reaches a mean Dice of 0.9002 (WT 0.9579, TC 0.8393, ET 0.9033), and 0.893 } 0.005 across repeated runs. Paired ablation tests confirm the contribution of graph connectivity and, for whole-tumor and enhancing-tumor Dice, of attention. The model has 1.36 M parameters; its network forward pass takes 0.10 s on a 16 GB laptop and uses about 1.5 GB of memory, although end-to-end processing (about 32 s per volume) is dominated by supervoxel generation. Post-hoc gradient-based saliency provides regional, modality-level attributions whose clinical value remains to be validated.