Knowledge distillation for glioma diagnosis: ResNet-based deep learning for WHO grade stratification with real-world deployment optimization.

Abstract

2083 Background: Gliomas are the most common primary CNS malignancies. WHO 2021 classification stratifies diffuse gliomas into molecularly-defined grades with vastly different treatment pathways: grades 2–3 (median survival 7–10 years) versus grade 4 glioblastoma (median survival 15 months untreated). Accurate preoperative grading is critical for surgical planning and treatment intensity. However, 5–8% of gliomas are initially misdiagnosed as meningiomas or pituitary adenomas, delaying definitive treatment by 3–6 months. Automated systems achieving high accuracy while remaining deployable in resource-limited settings could substantially reduce diagnostic error. Methods: Retrospective cohort: 2,840 brain MRI studies (1,400 gliomas: 700 grade 2–3, 700 grade 4; 900 meningiomas; 540 pituitary adenomas) from six continents. Ground truth used WHO 2021 classification; expert neuroradiologist consensus confirmed diagnoses. ResNet152 (60.2M parameters; 224×224 input; 28.6 GFLOPs) optimized for multi-sequence fusion (T1, T2, FLAIR, post-contrast) served as reference. ResNet18 (11.7M parameters; 224×224 input; 1.8 GFLOPs; 81% parameter reduction) underwent structured knowledge distillation—soft probability targets from ResNet152 with temperature-scaled softmax and regularized cross-entropy loss. Model performance evaluated on accuracy, sensitivity, specificity, F1-score, AUROC. Differential diagnostic accuracy assessed. Deployed globally; 52 neuroradiologists evaluated across six continents. Results: Glioma grading: ResNet18 achieved 96.9% accuracy for grade 2–3 vs. grade 4 (sensitivity 97.8%, specificity 96.0%, AUROC 0.991). Differential diagnosis: 95.1% for glioma vs. meningioma; 96.0% for glioma vs. pituitary. External validation confirmed consistent performance (93–97% grading; 94–96% differential diagnosis). Deployment advantage: ResNet18 inference time 38ms/image (ResNet152: 320ms), enabling real-time clinical integration on standard hospital GPU servers. Global deployment: 94.6% of neuroradiologists rated the system clinically valuable. Knowledge distillation preserved 99.2% of teacher diagnostic accuracy while reducing computational cost by 93.7%. Conclusions: Knowledge distillation enables ResNet18 to achieve ResNet152-comparable diagnostic accuracy for glioma grading and differential diagnosis while reducing parameters by 81% and computational cost by 94%. This framework solves critical deployment barriers in resource-limited healthcare systems. ResNet18's efficiency enables real-time inference on standard hardware, rapid diagnostic turnaround, and scalable global implementation. This system merits prospective evaluation to reduce time-to-diagnosis and eliminate glioma misclassification worldwide.