MULTIMODAL DEEP LEARNING FOR BRAIN TUMOR CLASSIFICATION: INTEGRATING CLINICAL, HISTOPATHOLOGICAL, AND DEMOGRAPHIC DATA WITHIN A UNIFIED FRAMEWORK

Abstract

Brain Tumor classification is still difficult because the decisions are made based on sporadic and heterogeneous information which is often analyzed separately. The study introduces a single, integrated multimodal deep-learning framework that combines histopathological images and clinical and demographic information to create more accurate, robust, and interpretable brain tumor classification. The proposed method uses a deep visual encoder to capture the morphological characteristics of tumors in the histopathological image and a specifically designed neural network to capture other patterns from the structured clinical and demographic variables. These modality-specific representations are fused together using an attention-based fusion mechanism, in order to obtain a global patient-level representation for tumor classification. The framework is compared to unimodal and conventional machine-learning baselines using class-sensitive discrimination and calibration metrics, and class-sensitive robustness metrics; and the contribution of each data modality is quantified through ablation studies. The use of EAI techniques to find influential tissue regions and patient characteristics is included, and subgroup analyses are performed to evaluate the demographic fairness and generalizability. The proposed framework combines pathological morphology with patient clinical context to build a more comprehensive and clinically relevant decision-support system for brain tumor diagnosis and set the stage for a scalable precision neuro-oncology framework.