Morphology-based molecular classification of spinal cord ependymomas using deep neural networks

Area of Science:

• Neuro-oncology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Spinal cord ependymomas (SP-EPN, MPE) have distinct molecular subtypes based on DNA methylation.
• Histomorphological diagnoses often mismatch molecular classification, impacting clinical management and prognosis.
• Accurate subtyping is crucial for effective treatment strategies.

Purpose of the Study:

• To develop and validate deep neural networks for predicting DNA methylation classes of spinal cord ependymomas from H&E stained whole-slide images.
• To utilize explainable AI to identify and quantify morphological features associated with molecular subtypes.
• To improve the consistency of histology-based diagnoses with molecular profiling.

Main Methods:

• A cohort of 139 molecularly characterized spinal cord ependymomas was analyzed.
• Self-supervised and weakly-supervised deep neural networks were employed for classification.
• Attention analysis and machine learning identified and quantified morphological features correlating with molecular types.

Main Results:

• The best-performing model achieved 98% test accuracy in predicting DNA methylation class from whole-slide images.
• Self-supervised learning outperformed pretrained networks (86% accuracy).
• AI predictions showed higher concordance with molecular classification (98%) than neuropathologists (83%).

Conclusions:

• Deep neural networks can accurately predict spinal cord ependymoma molecular subtypes using standard histology images.
• This AI-driven approach offers a potential supplementary tool for integrated diagnostics and standardization, particularly in resource-limited settings.
• Identified morphological features enhance understanding and consistency in ependymoma classification.