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

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Radiotherapy planning

Background:

• Auto-segmentation variability impacts clinical practice due to limitations of global similarity metrics.
• Anisotropic and localized contour disagreements are difficult to assess with current methods.

Purpose of the Study:

• To develop Deformable Point Cloud Registration-Based Bidirectional Local Distance (DPCR-BLD) for evaluating local auto-segmentation disagreements.
• To provide a systematic methodology for quantifying localized differences between auto-generated and clinical contours.

Main Methods:

• Quantified local differences using Bidirectional Local Distance (BLD) on reference and test structure point clouds.
• Employed deformable registration (coherent-point-drift) to propagate local discrepancies across datasets.
• Validated on 73 structures across four treatment sites for 1785 patients, developing an automatic outlier detection tool.

Main Results:

• DPCR-BLD effectively reveals systematic local disagreements and their magnitude in auto-segmentation.
• Demonstrated specific over- and under-contouring examples (e.g., brainstem).
• Identified prostate, seminal vesicle, and brainstem as organs with the most significant edits, with an effective outlier detection tool.

Conclusions:

• DPCR-BLD spatially identifies local contour differences between auto-generated and reference contours.
• The method facilitates contour outlier detection, aiding in quality assurance.
• Further prospective studies are needed to confirm the clinical utility of DPCR-BLD for AI-generated contour edits.