ResNets in the diagnosis of gliomas: Knowledge distillation from ResNet101 to ResNet18 for accurate and deployable WHO grade stratification.

Abstract

e14093 Background: Gliomas are the most common primary malignancies of the central nervous system and remain a leading cause of cancer-related mortality in young adults. The WHO 2021 classification stratifies gliomas into grades with markedly different prognoses, underscoring the importance of accurate, timely grading on magnetic resonance imaging (MRI). Despite advances in deep learning, many high-performing models are computationally intensive, limiting translation into routine clinical practice, particularly in resource-limited settings. Knowledge distillation offers a strategy to transfer diagnostic capability from high-capacity teacher networks to lightweight student models while preserving performance. We evaluated a ResNet-based distillation framework for glioma diagnosis and WHO grade stratification. Methods: We conducted a retrospective multicontinental analysis of 2,840 brain MRI studies, including gliomas (grades 2–3 and grade 4), meningiomas, and pituitary adenomas. Ground truth was established using WHO 2021 criteria with expert neuroradiologist consensus. A ResNet101 teacher model, leveraging deep residual connections for hierarchical feature learning across MRI sequences, was trained for glioma grading and differential diagnosis. A ResNet18 student model was trained using temperature-scaled knowledge distillation with regularized cross-entropy loss to inherit the teacher’s probabilistic decision structure while reducing computational complexity. Model performance was evaluated using accuracy, sensitivity, specificity, F1 score, and AUROC. Results: The distilled ResNet18 achieved high diagnostic performance for glioma grade stratification, with accuracy exceeding 96% and AUROC greater than 0.99 for distinguishing grade 2–3 from grade 4 disease. Differential diagnosis accuracy exceeded 95% for gliomas versus meningiomas and pituitary adenomas. Knowledge distillation reduced model parameters by over 80% and substantially decreased inference time, enabling real-time analysis on standard hardware. Neuroradiologist reviewers reported high clinical utility for diagnostic support. Conclusions: Knowledge distillation enables a lightweight ResNet18 model to achieve near–ResNet101 diagnostic performance for glioma diagnosis and WHO grade stratification while markedly reducing computational burden. This approach addresses key deployment barriers and supports scalable, real-time AI assistance in neuro-oncology imaging. Prospective validation is warranted to assess impact on diagnostic accuracy and time-to-treatment.