Abstract

Background: CNS tumors remain among the most frequently discordant pathologic diagnoses in the field of pediatrics. In this study, we examined neuropathologist-guided deep learning strategies towards automation of tumor histology diagnosis targeting the three most common pediatric Posterior Fossa (PF) tumors.

Methods: A retrospective chart review identified 252 pediatric patients with histologically confirmed PF Pilocytic Astrocytoma (PA); Ependymoma (EP); medulloblastoma (MB) across two independent institutions: Site 1: PA(n=87); EP(n=42); MB(n=50); Site 2: PA(n=36); EP(n=9); MB(n=28). The dataset comprised images of tumor-relevant regions captured by neuropathologists while viewing histology slides at 20 × magnification at the microscope. A Resnet-18 architecture was used to develop a 2D deep learning models and to assess model generalization across the two sites. Holdout test set was used to assess each of the model performance.

Results: Model trained exclusively on Site 1 cohort, achieved an accuracy of 0.75 and a F1 score of 0.61 on test set from Site 1; and an accuracy of 0.89 and F1 score of 0.77 on Site 2. Fine tuning on a subset of cohort from Site 2 did not significantly improve model performance.

Conclusion: We demonstrate a potential role implementing AI for histologic diagnosis of the three most common pediatric PF tumors that can generalize across centres. Further, we identify feasibility of AI learning that uses histology images captured by neuropathologists at the microscope and thereby incorporate expert human behavior. Future study could examine AI model developments that use tumor segmentations of histology slides in comparison to expert pathologist-guided image capture as forms of tumor labels.

Keywords: Central Nervous System (CNS); Posterior fossa; medulloblastoma; Pilocytic astrocytoma; Ependymoma