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

PedDx achieves 81.4% classification accuracy for pediatric posterior fossa tumor diagnosis on external validation using T2-weighted MRI, with non-abstained accuracy rising to 88.7%.

• Why it matters: Diagnostic AI needs evaluation that goes beyond accuracy to reveal uncertainty, failure modes, and robustness under distribution shifts for clinical translation.
• What they did: The segmentation-free framework uses pretrained 3DINO representations, calibrated predictions, and confidence-based abstention, then tests across discovery, replication, external validation, 236 OOD cases, and controlled imaging perturbations.
• The result: Independent external agentic evaluation on 40 cases yields 92.5% accuracy, improving to 97.3% among non-abstained cases, while OOD and perturbation analyses expose degradation and acquisition/alignment vulnerabilities.

AI summary of the abstract · may contain errors · not medical advice · read the paper.