Deep Learning-Based CT Segmentation of a Non-radiopaque Hydrogel Rectal Spacer in Prostate Radiotherapy

Introduction: The study evaluated the feasibility and geometric accuracy of CT-based automatic segmentation of a non-radiopaque hydrogel rectal spacer using a commercial deep learning platform (OncoStudio) and determined whether accurate spacer delineation can be achieved without MRI while preserving the contouring quality of adjacent organs. Methods: This retrospective study included 21 patients with localized prostate cancer who underwent external beam radiotherapy with a non-radiopaque hydrogel rectal spacer (SpaceOAR; Boston Scientific). CT-based automatic segmentation of the rectal spacer, prostate, rectum, and bladder was performed on the planning CT using OncoStudio (version 2.0; Oncosoft Inc.). Reference contours were manually delineated on the planning CT with reference to co-registered T2-weighted MRI. Geometric accuracy was assessed using the Dice similarity coefficient (DSC), 95th-percentile Hausdorff distance (HD₉₅), and mean surface distance (MSD). Results: The mean DSC for the rectal spacer was 0.787 ± 0.073, with a mean HD₉₅ of 4.03 ± 1.83 mm and a mean MSD of 1.18 ± 0.63 mm. The adjacent organs showed high geometric agreement, with mean DSC values of 0.876, 0.890, and 0.963 for the prostate, rectum, and bladder, respectively. No systematic over- or under-segmentation was observed, and the MSD for all structures was below 2.5 mm. Conclusions: CT-based automatic segmentation using OncoStudio provided clinically acceptable boundary delineation of a non-radiopaque hydrogel rectal spacer without requiring MRI, while maintaining the contouring quality of the surrounding organ. These findings suggest that CT-only automatic contouring can reduce spacer-specific MRI use and associated registration uncertainty in prostate radiotherapy.