Adaptive Vision–Language Transformer for Multimodal CNS Tumor Diagnosis

<b>Objectives</b>: Correctly identifying Central Nervous System (CNS) tumors through MRI is complicated by utilization of divergent MRI acquisition protocols, unequal tumor morphology, and a difficulty in systematically combining imaging with clinical information. This study presents the Adaptive Vision-Language Transformer (AVLT), a multimodal diagnostic infrastructure designed to integrate multi-sequence MRI with clinical descriptions while improving robustness and interpretability to domain shifts. <b>Methods</b>: AVLT integrates the MRI sequence (T1, T1c, T2, FLAIR) and clinical note text in a joint process using normalized cross-attention to establish association of visual patch embeddings with clinical token representations. An Adaptive Normalization Module (ANM) functions to mitigate distribution shift across datasets by adapting the statistics of domain-specific features. Auxiliary semantic and alignment losses were incorporated to enhance stability of multimodal fusion. <b>Results</b>: On all datasets, AVLT provided superior classification accuracy relative to CNN-, transformer-, radiogenomic-, and multimodal fusion-based models. The AVLT model accuracy was 84.6% on BraTS (OS), 92.4% on TCGA-GBM/LGG, 89.5% on REMBRANDT, and 90.8% on GLASS. AvLT AUC values are at least above 90 for all domains. <b>Conclusions</b>: AVLT provides a reliable, generalizable, and clinically interpretable method for accurate diagnosis of CNS tumors.