A functionally guided fusion Vision Transformer for predicting IDH status in gliomas: a multicenter study with external validation and incomplete multimodal evaluation.

Abstract

Background: Accurate preoperative prediction of isocitrate dehydrogenase (IDH) genotype in gliomas is crucial for treatment planning and prognostic evaluation. However, variability across imaging modalities and centers limits model generalization in clinical practice.

Purpose: We aimed to develop and evaluate a functionally guided fusion Vision Transformer (FGF-ViT) network for IDH genotype prediction in gliomas and to assess its generalization across multicenter datasets and incomplete multimodal inputs.

Methods: This retrospective multicenter study involved glioma patients from multiple institutions. In step 1, four FGF-ViT networks were constructed using different modality combinations (conventional MRI [cMRI]; MRI + diffusion-weighted imaging [DWI]; cMRI + perfusion-weighted imaging [PWI]; cMRI + DWI + PWI), trained on a primary cohort, and tested on an independent external validation set. Step 2 evaluated model generalization on additional multicenter datasets with variable modality availability. Models fused cMRI, DWI, and DSC-PWI features via transformer attention. Performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity.

Results: The FGF-ViT achieved robust IDH prediction with an AUC of 0.822 (95% CI: 0.666-0.977) in the independent external validation cohort. Its performance remained stable even with one missing functional modality.

Conclusion: The proposed FGF-ViT provides a clinically relevant multimodal imaging and generalizable framework for preoperative IDH genotype prediction in gliomas, enabling reliable application across centers and incomplete multimodal conditions.