Use of knowledge-distilled EfficientNet to enable rapid, scalable glioma diagnosis.

Abstract

2040 Background: Accurate preoperative classification of gliomas on MRI according to WHO 2021 integrated diagnosis is essential for surgical planning and treatment selection, yet conventional interpretation remains vulnerable to inter-observer variability. Deep learning shows promise for brain tumor classification but computational demands restrict deployment in resource-limited settings. Knowledge distillation, transferring discriminative knowledge from computationally intensive "mentor" networks to efficient "mentee" models, enables scalable AI systems maintaining diagnostic accuracy while enabling global translation. Methods: We analyzed 2,840 brain MRI studies sourced from six continents comprising gliomas (n=1,400; 700 WHO grade 2–3, 700 grade 4), meningiomas (n=900), and pituitary adenomas (n=540). Ground truth followed WHO 2021 integrated diagnosis with expert neuroradiologist consensus. Data were partitioned at the patient level with site-held-out external testing. A high-capacity EfficientNetB7 mentor model was developed using multi-sequence MRI inputs (T1, T2, FLAIR, post-contrast T1) and distilled into lightweight EfficientNetB0. Knowledge distillation preserved discriminative performance while enabling CPU deployment without GPU acceleration. Training: input 224×224 pixels, Adam optimizer (lr=0.001), Kullback-Leibler divergence loss (temperature T=4), twelve epochs. Performance metrics included accuracy, sensitivity, specificity, F1-score, and AUROC. Independent clinical feasibility involved 52 board-certified neuroradiologists across 11 geopolitical regions. Results: For WHO grade 2–3 versus grade 4 stratification, distilled EfficientNetB0 achieved 97.2% accuracy, 98.1% sensitivity, 96.3% specificity, and AUROC 0.993. Differential diagnosis accuracy: 95.3% (glioma vs. meningioma), 96.1% (glioma vs. pituitary). Site-held-out external validation showed consistent performance (93–97% grading, 94–96% differential). Mean inference time: 0.19 seconds on CPU. Distilled model: 3.9 million parameters (97% reduction), 2.1 billion FLOPs (94% reduction). Neuroradiologist assessment: 94.7% rated system valuable for diagnostic confidence and surgical triage. Inter-rater agreement: 96.8% (κ=0.951, 95% CI 0.941–0.961). Conclusions: Knowledge-distilled EfficientNetB0 achieves 97.2% accuracy for glioma grading and 95–96% for differential diagnosis while reducing parameters by 97% and enabling CPU deployment. Global validation across six continents and clinical assessment by 52 neuroradiologists establish real-world feasibility. The architecture supports translation as a decision-support tool for preoperative glioma stratification, particularly in resource-constrained settings. Prospective clinical trials evaluating impact on surgical decision-making are warranted.