Location smoothed Bayesian additive regression trees: a method for interpretable and robust quality assurance of organ contours in radiotherapy treatment planning

Abstract Deep learning techniques for image segmentation are increasingly used in automating anatomical structure delineation in medical images for radiation treatment planning. Given the critical role these contours play in guiding radiotherapy, it is crucial to flag errors before planning, necessitating robust quality assurance methods for the clinical adoption of automated contours. To address this challenge, we introduce location smoothed Bayesian additive regression trees (lsBART), a novel Bayesian tree-based model for nonparametric scalar on function regression. Our proposed method can identify both relevant functions and important regions within those functions, enabling interpretable, and sparse solutions. We benchmark lsBART on a simulated regression setting with multiple functional predictors, where it achieves a lower root mean squared error than existing alternative methods. In our real data application to identifying errors in kidney contours, we attained a cross-validated area under the curve of 0.905 for detecting unacceptable contours. Using Shapley values, we provide guidance on aspects of the contour in specific regions that led to the contour being flagged, indicating our method’s potential clinical utility.