GLORIA AI fuses histology, genomics, and MRI for superior glioma grading
Trimodal alignment using a novel Gramian contrastive loss outperforms bimodal methods by double digits
GLORIA (GLioma Omics-Radiology-hIstopathology Alignment) processes whole-slide image regions, mRNA expression data, and 3D MRI volumes through separate encoders, projecting them into a shared latent space. The key innovation is a Gramian contrastive loss that measures the volume spanned by the three modality embeddings, forcing tighter alignment than pairwise methods. A cross-modal gating module then fuses the aligned representations for joint optimization on three-class glioma grading and overall survival prediction.
Evaluated on 132 matched patients from TCGA-GBM/LGG and BraTS21, GLORIA consistently outperformed the strongest bimodal baseline (WSI-mRNA) on all evaluated metrics. This approach captures complementary information from histopathology (tissue morphology), transcriptomics (molecular activity), and radiology (tumor extent), enabling more robust and interpretable tumor characterization. The work points toward a future where AI-driven diagnosis integrates multiple data sources seamlessly.
- Combines three medical data types: whole-slide histopathology images, mRNA expression profiles, and 3D MRI volumes
- Novel Gramian contrastive loss aligns multimodal embeddings by measuring the volume they span in latent space
- Achieves state-of-the-art results on TCGA/BraTS21 cohort (132 patients), surpassing bimodal WSI-mRNA baselines
Why It Matters
Enables more accurate and comprehensive glioma prognosis by fusing radiology, genomics, and pathology into a single AI model