Quantitative Evaluation of Auto Contouring Algorithm Performance on Hypersight Versus Conventional CBCT
Abstract
Purpose
Accurate auto-contouring is essential for efficient prostate radiotherapy, particularly in image-guided and adaptive workflows, where contour quality influences clinical decision-making. This study quantitatively evaluates Radformation auto-contouring performance on planning CT (pCT), HyperSight CBCT (HS-CBCT), and traditional CBCT (tCBCT) using drawn contours as ground truth.
Methods
This retrospective study included 10 prostate radiotherapy patients with planning CT (n=10), HS-CBCT (n=145), and traditional CBCT (n=82) acquired during treatment. Auto-contours for the prostate, bladder, rectum, and seminal vesicles were generated using Radformation. Ground-truth contours were manually delineated/reviewed by experienced radiation oncologists on all scans. Agreement between auto-generated and manual contours was quantified using Dice Similarity Coefficient (DSC), 95th percentile Hausdorff distance (HD95), Structural Similarity Index Measure (SSIM), and added path length (APL). The statistics presented here focused on contours relevant for treatment planning, and thus only included slices ±5cm from the prostate. Analyses were stratified by imaging modality and anatomical structure. While full cohort analysis is ongoing, supporting data is included from four patients with bladder and rectum contours.
Results
Quantitative evaluation across all assessed metrics demonstrated the highest auto-contouring agreement on planning CT, with reduced accuracy observed on HyperSight CBCT and further degradation on traditional CBCT. Planning CT contours showed strong agreement with clinician-drawn ground truth, while HyperSight CBCT maintained reasonable geometric fidelity with increased variability. Traditional CBCT exhibited greater variability and reduced accuracy, most notably for bladder delineation. Qualitatively in these scans, auto-contours frequently extended into adjacent bowel structures or incompletely contoured the bladder. Such auto-contouring errors were not observed on planning CT or HyperSight CBCT images.
Conclusion
Auto-contouring accuracy follows a consistent hierarchy: planning CT performs best, HyperSight CBCT provides comparable but slightly reduced performance, and traditional CBCT shows the greatest limitations, particularly for bladder delineation.