MDB-02. Automated detection of histopathological features using a linear classifier model in anaplastic medulloblastoma

Background Large Cell/Anaplastic medulloblastoma is the most aggressive histologic subtype and is associated with poor outcomes. Hallmark features of anaplasia include enlarged, pleomorphic nuclei and increased mitotic activity, but the absence of a standardized classification system has led to inconsistent application of histological criteria. This inconsistency hinders reproducible evaluation, underscoring the need for innovative, objective approaches to histopathological analysis. Methods In this study, we developed a computational pipeline that leverages a combination of specialized software including QuPath (v0.5.1), PyTorch, and Scikit-learn to curate regions of interests (ROI), conduct transfer-learning feature extraction, and train a classification model. The pipeline automates ROI extraction, splits ROIs into tiles, confirms tissue presence in each tile, then extracts features and trains a linear classifier. We obtained 42 whole slide images (WSIs) representing anaplastic medulloblastoma from CHLA. Samples contained varying extents and severity of anaplasia, as well as fragments of adjacent cerebellar cortex. Regions containing medulloblastoma with either anaplastic (n = 68,138) or classic (n = 42,174) histology were manually annotated per WHO definitions then converted into 600 × 600 pixel tiles (∼228² μm2). Transfer-learning feature extraction was conducted using a foundational model (UNI via DINOv2) to extract 1024 features per tile. Data were split across WSIs to prevent data leakage using a 70/30 split for training and validation. Finally, classifier predictions were graphically mapped back to WSI, providing an overview of the distribution of anaplasia within each tumor sample. Results Our ROI classification model detected anaplasia in the test set with an accuracy of 96% and achieved an AUROC of 0.999. Conclusion This study provides proof-of-concept for leveraging deep learning to differentiate between anaplastic and classic morphologies in pediatric medulloblastoma, laying the groundwork for standardizing the diagnosis of this clinically important histological variant.