PedDx: An End-to-End Deep Learning Framework for Pediatric Posterior Fossa Tumor Diagnosis with Agentic AI–Enabled Independent External Evaluation

Reliable clinical translation of diagnostic artificial intelligence (AI) requires evaluation beyond classification accuracy. We developed PedDx, a segmentation-free framework for pediatric posterior fossa tumor diagnosis using T2-weighted MRI, pretrained 3DINO representations, calibrated predictions, and confidence-based abstention. Evaluation included 492 discovery, 114 replication, 129 external validation, and 40 independently tested cases, alongside 236 out-of-distribution (OOD) cases and controlled imaging perturbations. Classification accuracy was 79.8% in replication and 81.4% in external validation, increasing to 90.5% and 88.7%, respectively, among non-abstained cases. OOD analyses revealed performance degradation and confidently incorrect predictions, while perturbation analyses identified acquisition and alignment vulnerabilities. An external institution independently executed the complete agentic workflow, successfully generating predictions and structured quality-control reports for all 40 cases, with classification accuracy of 92.5%, increasing to 97.3% among non-abstained cases. PedDx demonstrates how uncertainty assessment, failure-mode characterization, external validation, and independently reproducible, traceable reporting can collectively inform the reliability and limitations of diagnostic AI systems.