Graph-Aided Diffusion Models for Precision Segmentation of Challenging Tumor Areas in MRI

Motivation: Building on previous work, this study aims to improve glioblastoma segmentation by focusing on complex tumor sub-regions, particularly in low-contrast areas. Goal(s): Our goal is to enhance segmentation precision across tumor regions by combining Graph Neural Networks (GNNs) with diffusion models for stable multimodal MRI performance. Approach: GNNs capture global and local features, creating coarse segmentation masks that guide U-Net performance on smaller regions. These representations are reprojected into volumetric space, with boundary penalties to refine accuracy. Results: Our approach improves segmentation in challenging sub-regions, balancing performance on smaller regions and the overall tumor, supporting more precise glioblastoma treatment planning. Impact: This study builds on prior work by combining diffusion models with graph neural networks to enhance brain tumor segmentation. By utilizing focused graph representations, the proposed method improves precision particularly within the tumor core and smaller subregions.