A novel physics-guided graph transformer (PHY-GT) for multimodal brain tumor classification

Abstract Background Accurate classification of brain tumors from magnetic resonance imaging (MRI) is critical for effective clinical diagnosis and treatment planning. Although deep learning models such as convolutional neural networks and vision transformers have demonstrated strong performance, challenges remain in handling imaging heterogeneity, limited annotated data, and inadequate modeling of tumor structural characteristics. The aim of this study is to develop and evaluate a novel Physics-Guided Graph Transformer (PHY-GT) framework that integrates transformer-based contextual learning, graph-based structural reasoning, and physics-guided regularization for automated brain tumor classification from MRI images. Results The proposed PHY-GT integrates convolutional feature extraction, transformer-based global contextual modeling, graph-based tumor topology representation, and physics-guided learning into a unified architecture. The model processes multimodal MRI sequences (T1, T1ce, T2, and FLAIR) and incorporates a graph attention mechanism to capture intratumoral spatial relationships. A physics-guided regularization strategy enforces consistency across imaging modalities and is designed to promote stable feature learning across heterogeneous MRI representations. Experimental evaluation on a publicly available brain tumor MRI dataset demonstrated improved performance compared to state-of-the-art models, achieving an accuracy of 94.5%, precision of 95.2%, recall of 95.5%, and F1 score of 95.3%. Ablation studies confirm the contribution of each component, including physics-guided learning and graph-based modeling, to overall performance improvements. Conclusions The PHY-GT framework provides a proof-of-concept framework for automated brain tumor classification using MRI data. By effectively combining multimodal feature fusion, structural reasoning, and physics-informed learning, the proposed model demonstrates improved classification performance on the evaluated dataset. The proposed PHY-GT demonstrates promising experimental performance for MRI-based brain tumor classification. However, the present study should be interpreted as a proof-of-concept evaluation based on a single publicly available dataset, and further external multicenter validation is required before broader clinical applicability can be established.