Deep Learning‐Based Multimodal Fusion of Whole‐Slide Images and RNA Sequencing Identifies Survival‐Relevant Glioblastoma Clusters

ABSTRACT

Glioblastoma is profoundly heterogeneous, and single‐modality analyses often miss prognostically relevant structure. We introduce a transparent, end‐to‐end workflow that fuses available whole‐slide histology and RNA‐seq to discover clinically meaningful glioblastoma subgroups using an unsupervised learning model after feature extraction. Haematoxylin–eosin slides are tiled, tissue‐screened and stain‐normalised; tiles are embedded with a pretrained ResNet‐50 to yield 2048‐dimensional features, averaged per patient and compressed to 30‐D by an autoencoder. In parallel, RNA‐seq (~48 k genes) undergoes low‐variance filtering and normalisation, then a second autoencoder produces a 30‐D transcriptomic embedding. The two 30‐D representations are concatenated into a 60‐D fused vector, robustly scaled and refined with PCA (≈98% variance retained). Across K ‐means, Gaussian mixture models and Agglomerative clustering ( k  = 2–20), Agglomerative k  = 2 was decisively best (mean silhouette ≈0.53), yielding clusters of 150 and 8 patients (survival subset 147 and 8). Survival separation was substantial (median 454 vs. 138 days; log‐rank p  = 0.0096). In Cox models, the poorer‐prognosis cluster showed increased risk (HR ≈ 2.70), which remained significant after age adjustment (HR = 2.15, 95% CI 1.04–4.46; age per year HR = 1.02, 95% CI 1.01–1.04). Attribution and consensus analyses yielded compact, interpretable gene sets (22 shared; 8 per cluster), including markers associated with NOTCH/γ‐secretase and oxidative phosphorylation. These findings nominate biologically plausible hypotheses for future validation rather than immediate treatment‐selection rules. Overall, this study demonstrates that auditable late fusion of histology and transcriptomics, built from routine data, can identify survival‐associated glioblastoma subgroups and provides a hypothesis‐generating framework for prospective, harmonised, multi‐centre validation.