Leveraging Graph Neural Networks in Transferring Multimodal Knowledge for Unimodal Segmentation

Accurate segmentation of brain tumors in MRI scans is crucial for effective treatment. Multimodal MRI, including FLAIR, Tlce, T1, and T2 images, provides valuable infor-mation for tumor delineation. However, challenges such data corruption and varying protocols can lead to missing modal-ities, affecting segmentation accuracy. Recent methods at-tempt to transfer knowledge from multimodal teacher mod-els to unimodal student models, but often struggle to capture complex structural relationships within the data. This paper introduces a novel approach called Graph Neural Network-based Knowledge Distillation (GKD), which uses Graph Neu-ral Networks to create a comprehensive graph connecting in-termediate layers of both teacher and student networks. This enables the student network to better learn essential features and their interactions. Evaluations on the BraTS2018 and BraTS2020 datasets demonstrate that GKD significantly en-hances brain tumor segmentation accuracy by deepening the connection between teacher and student insights. The code is publicly available at https://github.com/T-Y-Liu/GKD.