Biostatistics Seminar: Pediatric Deep Learning and Statistical Quality Assurance

About this event

Building Trustworthy Autosegmentation: Pediatric Deep Learning and Statistical Quality Assurance

Zachary Wooten, Ph.D. St. Jude Children’s Research Hospital

In medical image autosegmentation, a segmentation model can perform well on average and still miss part of an organ in an individual patient's scan. In radiation therapy, where organ contours guide treatment planning, these errors matter. While autosegmentation can reduce the time spent on manual contouring, clinical use also depends on our ability to identify contours that need closer review. In this seminar, I will discuss our work developing a pediatric-trained deep learning pipeline for craniospinal irradiation and evaluating it against atlas-based and commercial AI approaches. In a study of 211 pediatric CT scans, the in-house model achieved higher average segmentation accuracy than both alternatives. Beyond overall accuracy, we used mixed-effects models to investigate differences across organs and patient ages, along with learning curves and dosimetric analyses to better understand model performance. Even with these improvements, identifying individual segmentation errors remains a challenge. I will discuss two statistical approaches for quality assurance of autosegmentations. The first uses kernel density estimation of CT intensity and shape features to flag unusual contours without requiring labeled examples of errors. The second uses a supervised approach, location-smoothed Bayesian additive regression trees (lsBART), to predict contour acceptability from functional shape predictors and highlight anatomical regions that influence those predictions. In a separate contour study, lsBART achieved a cross-validated area under the curve of 0.905 in error detection. I will end the talk by discussing how these methods could be brought together into a clinical workflow, and the challenges that remain. More broadly, this work highlights why statistical methods remain essential even after an AI model has been developed, from evaluating performance across patient populations to understanding where models fail and how those failures can be identified.