Deformable Point Cloud Registration-Based Bidirectional Local Distance (DPCR-BLD): A Methodology for Systematic...

Variability in autosegmentation can arise from the training data, leading to disagreements with clinical practice. The anisotropic and localized nature of disagreements between 2 contours makes them challenging to evaluate using common metrics, which only provide insights on overall and global similarity. This study aims to develop Deformable Point Cloud Registration-Based Bidirectional Local Distance (DPCR-BLD), a methodology to systematically evaluate local disagreements in autosegmentation.

Given a reference (clinically approved) and test (autogenerated) structure data set, BLD was employed to quantify the local differences between test and reference structure point clouds. Using a validated reference contour as the template contour, each reference contour point cloud with assigned BLDs was deformably registered via the coherent-point-drift algorithm to propagate local discrepancies across the data set. The proposed methodology was validated on 2 independent retrospective data sets including 73 structures across 4 common treatment sites for 1785 patients. An automatic outlier detection tool was also developed by determining the number of points falling outside defined thresholds.

The DPCR-BLD methodology can reveal systematic local disagreements in different regions, offering insights into the magnitude and variability of how clinicians edit autocontours in clinical practice. For instance, the brainstem autosegmentation tends to overcontour the central superior region by 1 mm, whereas undercontouring the peripheral superior region by 1 mm. The automatic detection tool was able to detect statistical outlier, with the top 3 organs flagged as major edit are prostate (32.9%, n = 51), seminal vesicle (23.5%, n = 36), and brainstem (18.7%, n = 72). Template contour selection has minimal impact on results, once it adequately represents the organ morphology.

DPCR-BLD provides a mechanism to spatially identify local contour differences between 2 contour sets. Also, we demonstrate how this method could be used for contour outlier detection. Further work is needed to demonstrate the clinical utility of this tool in prospective evaluation of edits to artificial intelligence-generated contours.