PATH-06. IMAGE-BASED MACHINE LEARNING CLASSIFIER FOR PEDIATRIC POSTERIOR FOSSA TUMOR HISTOPATHOLOGY. (4th December 2020)

• Record Type: Journal Article
• Authors: Tam, Lydia; Bala, Wasif; Lavezo, Jonathan; Lummus, Seth; Vogel, Hannes; Yeom, Kristen
• Is Part Of: Neuro-oncology. Volume 22(2020)Supplement 3
• Page Start: iii425
• Page End: iii425
• Publication Date: 2020-12-04
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noaa222.642
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: BACKGROUND: Pediatric posterior fossa (PF) tumors can include astrocytomas, ependymomas, and medulloblastomas, all of which demonstrate unique histopathology. Whole slide image analyses can be time consuming and difficult. Therefore, we used machine learning to create a screenshot-based histopathology image classifier that can distinguish between types of pediatric PF tumors. METHODS: We took 179 histopathology slides from Stanford University, dated from 2008–2019: 87 astrocytomas, 42 ependymomas, and 50 medulloblastomas, per pathology report. Each slide was viewed under a microscope at 20x. Then, a screenshot was taken of the region of interest representative of principal slide pathology, confirmed by a trained neuropathologist. These screenshots were used to train Resnet-18 models pre-trained on the ImageNet dataset and modified to predict three classes. Various models with different hyperparameters were trained using a random hyperparameter search method. Trained models were evaluated using 5-fold cross-validation, assigning 20% of the dataset for validation with each evaluation. Qualitative analysis of model performance was assessed by creating Class Activation Map (CAM) representations of image predictions. RESULTS: The top performing Resnet-18 model achieved a cross-validation F1 of 0.967 on categorizing screenshots of tumor pathology into three types. Qualitative analysis using CAMs indicated the model was able to identify salient distinguishing features of each tumor type. CONCLUSIONS: We present a PF lesion classifier capable of distinguishing between astrocytomas, ependymomas, and medulloblastomas based on a histopathology screenshot. Given its ease of use, this tool has potential as an educational tool in an academic setting.