BIOM-94. MULTIMODAL EXPLAINABLE AI FOR MOLECULAR SUBTYPING OF GLIOMAS USING RADIOLOGY AND PATHOLOGY INTEGRATION

Le résumé fourni par la source

Abstract Accurate molecular subtyping of gliomas is essential for guiding integrated diagnosis, predicting prognosis, and informing clinical decision-making in neuro-oncology. While imaging and pathology each provide valuable perspectives, unimodal approaches often fail to capture the full biological and spatial complexity of gliomas. In this study, we present a biologically interpretable, multimodal AI model (Mc) that integrates handcrafted features from multi-sequence MRI and digitized H&E-stained histopathology (MP) whole slide images (WSIs) to classify six clinically relevant molecular subtypes. Our cohort included 117 glioma patients with annotated molecular profiles per WHO 2021 classification. Of these, 50 patients had both pre-operative MRI and WSIs, enabling multimodal analysis. Tumor subregions were segmented on MRI using nnU-Net, and 110 radiomic features were extracted, capturing intensity, shape, and texture from enhancing, necrotic, and edematous regions. WSIs were annotated by pathologists and processed using HoVer-Net for nuclear segmentation, yielding 242 features characterizing nuclear morphology and spatial structure. The top five features from each modality were selected and concatenated into a unified multimodal feature set. Classification was performed using support vector machines and evaluated via cross-validation. The multimodal model Mc outperformed MRI- (MR) and pathology-only (MP) models across all subtypes, achieving AUCs of, IDH mutation: 0.822, CDKN2A/B homozygous deletion: 0.904, chromosome 7 gain: 0.848, chromosome 10 loss: 0.805, 1p/19q co-deletion: 0.803, and MGMT promoter methylation: 0.743. Feature attribution revealed biologically meaningful patterns: FLAIR-based edema features were predictive of 1p/19q co-deletion, while nuclear shape irregularities correlated with MGMT status. The use of interpretable, handcrafted features supports clinical trust and integration into decision-making workflows. This work demonstrates that fusing radiologic and histopathologic information enables more accurate, explainable, and clinically actionable glioma subtyping. The framework lays the foundation for real-world deployment of precision AI in neuro-oncology and will be further validated in multi-institutional, prospective studies.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.