Rule-aware reasoning visual language model for molecular prediction in adult-type diffuse glioma lacking contrast enhancement

Accurate preoperative prediction of isocitrate dehydrogenase(IDH) mutation and 1p/19q codeletion is crucial for treatment planning in adult-type diffuse glioma lacking contrast enhancement(ADG-LCE). Existing methods face a trade-off between interpretability and automation. We developed a visual language model (VLM) framework for interpretable molecular prediction in ADG-LCE using standard CT and MRI. This retrospective multi-center study included 873 pathologically confirmed ADG-L\CE patients. Molecular Inference-Guided Semantic Vision–Language Model(MIG-SemVLM) extracted five semantic imaging features (T2-FLAIR mismatch sign, cortical involvement, calcification, FLAIR homogeneity, and tumor border), which were integrated with age and ADC ratio via a rule-aware LLM following a simple scoring system. Model performance was evaluated by AUC, accuracy, F1-score, and balanced accuracy, with human-AI collaboration assessed in 300 cases. MIG-SemVLM achieved excellent semantic extraction performance across five features (AUC range: 0.9215–0.9536). The rule-aware LLM demonstrated strong molecular prediction on the test set: IDH mutation (AUC: 0.9349, accuracy: 0.8806) and 1p/19q codeletion (AUC: 0.9126, accuracy: 0.8538). AI assistance was associated with improved diagnostic accuracy among readers (IDH: +7.0%, 1p/19q: +7.0%) and reduced decision time by 56.2 s per case in the exploratory human–AI reader study. The proposed VLM framework with rule-aware reasoning provides accurate, interpretable molecular prediction in ADG-LCE, benefiting less experienced readers in clinical decisions.