Abstract
BACKGROUND: Pilocytic astrocytoma (PA) is the most frequent pediatric brain tumor, classified as CNS WHO grade 1. The 2021 WHO CNS tumor classification emphasizes integrated molecular diagnosis; accordingly, PAs are defined by key MAPK pathway alterations, most notably the KIAA1549-BRAF fusion. While this molecular stratification improves diagnostic accuracy, it can also belinked to technical issues and high turanaround time. In parallel, advances in digital histopathology and artificial intelligence (AI) are opening new avenues in neuro-oncology diagnostics. Recent deep learning models have automated brain tumor classification and even inferred molecular markers directly from hematoxylin and eosin slides. However, pediatric tumors remain underrepresented in computational pathology, partly due to a lack of large dedicated datasets. MATERIAL AND METHODS: We assembled a cohort of 70 pediatric PA cases. Of these, 32 harbored the KIAA1549-BRAF fusion. We developed a model based on a Vision Transformer (ViT) pretrained on histopathological data and fine-tuned using Low-Rank Adaptation (LoRA). The model employed a weakly supervised multiple instance learning framework: patches were extracted from each Whole Slides Imaging and aggregated via an attention mechanism for slide-level binary classification (fused vs. non-fused). Training used a pilot set of 6 cases: 2 for training, 2 for validation, and 2 for testing. RESULTS: In the preliminary phase, the fine-tuned ViT achieved 100% accuracy on the independent test set, correctly distinguishing between fused and non-fused PAs. The explainability framework provided clear visualizations, with attention maps focusing on classic PA histopathological structures such as piloid regions and Rosenthal fibers, aligning well with expert neuropathologist interpretation. CONCLUSION: This study introduces a novel pediatric-specific AI tool designed to predict molecular alterations in Pas directly from histological slides, aligned with the 2021 WHO CNS classification’s emphasis on integrated diagnostics. By fine-tuning a ViT with LoRA on a limited but curated dataset, we demonstrate that deep learning models can achieve high diagnostic precision with minimal data. Our findings highlight the importance of explainable AI in pediatric neuropathology and support the integration of AI into diagnostic workflows. Further validation on larger cohorts is ongoing to confirm the model’s generalizability and clinical applicability.