Machine learning based identification of anaplastic features in pediatric cancers

Background: Across cancers, histologic anaplasia is a poorly defined morphologic feature, signified by cytologic variability, often associated with relatively poorer outcomes, and requiring subjective judgment for definitive pathologist interpretation. Anaplasia is a key prognostic feature in two common pediatric embryonal malignancies, medulloblastoma (MB), occurring in the brain, and Wilms Tumor (WT), which arises from the kidney. Anaplastic histologic variants of these tumors are rare, comprising roughly 5-10% of all pediatric WT and 5-20% of MB. MB and WT each affect 500-600 patients per year in the US, and in both tumor types histologic anaplasia is a recognized feature of poor prognosis, associated with inferior overall survival.

Methods: We developed a computational pipeline incorporating image segmentation, foundational model-driven feature extraction, and hierarchical machine learning model training to classify regions of anaplasia among MB whole slide images (n=85, Aperio AT2 scanner). This pipeline was then applied to a WT cohort (n=41) to evaluate sensitivity and specificity for anaplasia classification.

Results: The model demonstrated robust performance classifying anaplastic and non-anaplastic 256x256 pixel tiles (AUROC 0.998, balanced accuracy 92.86%, weighted F1 0.968). Preliminary inference demonstrates detection of anaplastic regions with quantification of anaplastic burden across WT samples.

Conclusion: These findings support the generalizability of histologic anaplasia models across distinct tumor types and contexts. Given that anaplasia is rare yet prognostically significant in both MB and WT, cross-tumor models of this kind may offer meaningful clinical decision support for a diagnostically challenging feature.

Learning Objectives:

1. Recognize the prognostic significance of histologic anaplasia in pediatric cancers
2. Evaluate the performance of a whole slide image computational pipeline to identify a specific cytopathologic feature
3. Understand how predictive models can provide decision support for microscopic diagnoses in pediatric pathology