DEEP LEARNING APPLICATIONS IN PEDIATRIC ONCOLOGY: TUMOR DETECTION, CLASSIFICATION, AND TREATMENT RESPONSE ASSESSMENT

Abstract

Background: Pediatric cancers differ substantially from adult malignancies in biology, histology, and treatment sensitivity, necessitating specialized AI models. Objective: To evaluate deep learning model performance for pediatric tumor detection, histological classification, and treatment response prediction across major pediatric cancer types. Methods: A multicenter retrospective study analyzed imaging and molecular data from 4,312 pediatric oncology patients across seven children's cancer centers. Convolutional neural networks and graph neural networks were developed for tumor segmentation, WHO classification, and survival prediction. Results: Tumor segmentation achieved mean Dice similarity coefficient of 0.87 across brain tumors, Wilms tumors, and neuroblastoma. Histological classification of brain tumors matched WHO 2024 molecular classification in 89.3% of cases. Survival prediction models achieved C-index of 0.78 at 5-year follow-up. Conclusion: Deep learning demonstrates clinically meaningful performance in pediatric oncology applications, with particular strength in tumor segmentation and molecular classification prediction from histological imaging.