GMENet uses generative AI to diagnose glioma from incomplete MRI scans

A team of researchers (Pengfei Song et al.) has introduced GMENet, a novel deep learning framework designed to tackle a critical bottleneck in medical AI: missing MRI sequences. In clinical practice, imaging protocols vary across hospitals, leading to incomplete sequences that force many models to discard data. GMENet addresses this with two key innovations. First, a Cross-attention-based Gated Generation Module synthesizes missing sequence features from available ones using cross-attention and dynamic gating, with a cycle-consistency loss to preserve semantic integrity. Second, a Dynamically Weighted Experts Fusion Module performs mixture-of-experts interaction and confidence-aware fusion over both original and synthesized features, enabling robust multi-task prediction (e.g., molecular subtype and grade classification).

Evaluated on a multi-center cohort of 1,241 subjects from four in-house datasets and two public repositories, GMENet demonstrated exceptional performance. It expanded the pool of clinically usable training data by 97% compared to using only complete sequences. More importantly, it consistently outperformed state-of-the-art methods that had been trained exclusively on complete data, showing improved robustness under cross-center distribution shifts. The paper, accepted at IJCAI 2026, represents a significant step toward making AI-powered glioma diagnosis practical in real-world, heterogeneous clinical settings where incomplete data is the norm.