1074 - A Phase II Randomized Blinded Trial of AI-Assisted Versus Manual Contouring by New Learners in Ultrasound-Based Prostate HDR Brachytherapy
Presenter(s)
B. Thomsen1, N. Moldovan2, A. C. Smart3, P. L. Lee1, Y. H. Chen4, J. V. Mendoza1, C. Ciausu5, I. Buzurovic3, T. C. Harris3, D. Brivio3, C. E. Kehayias6, C. V. Guthier7, and M. T. King3; 1Department of Radiation Oncology, Mass General Brigham, Harvard Medical School, Boston, MA, 2Department of Radiation Oncology, Brigham and Women’s Hospital/Dana-Farber Cancer Institute, Boston, MA, 3Department of Radiation Oncology, Brigham and Women’s Hospital, Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA, 4Dana Farber Cancer Institute, Boston, MA, 5Artificial Intelligence in Medicine (AIM) Program, Mass General Brigham, Harvard Medical School, Boston, MA, 6Department of Radiation Oncology, Dana-Farber Cancer Institute/Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, 7Department of Radiation Oncology, Mass General Brigham/Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA
Purpose/Objective(s):
Contouring is one of the most clinically deployable applications of artificial intelligence (AI) in radiation oncology. Accurate contouring is essential for treatment planning but remains time-intensive, operator-dependent, and variable, particularly when adopting technically complex procedures. Despite strong performance in retrospective studies, prospective randomized data evaluating real-time clinical integration of AI are lacking. We conducted a prospective phase II randomized trial evaluating AI-assisted contouring within a live clinical workflow, using transrectal ultrasound (TRUS) -based prostate high-dose-rate (HDR) brachytherapy as a model of a technically demanding procedure. The hypothesis is that AI improves new learner contours with higher mean dice similarity coefficient versus manual contouring alone when compared with attending contours.Materials/Methods:
Patients receiving whole gland prostate HDR brachytherapy as monotherapy, boost, or salvage were eligible. New learners (residents, fellows, or attendings with <20 prior TRUS-based HDR cases) were randomized 1:1 to manual versus AI-assisted contouring using a previously validated U-Net algorithm, with stratification by learner type. For the manual group, learners contoured the target volume and organs at risk from scratch. For the AI-assisted group, learners edited AI-generated contours. A blinded, experienced brachytherapy attending then modified the learner contours for clinical treatment. The Dice similarity coefficient (DSC) was used to assess the volumetric similarity of learner contours against the final attending-modified contours used for treatment planning. The primary endpoint was the prostate DSC. Comparison between the two arms was done using t-test. Mixed effects model adjusting for learner stratification and incorporating intra-cluster correlation (correlation among cases done by the same new learner) was performed. The secondary endpoint was time from image acquisition to final contour approval. This trial was registered on ClinicalTrials.gov (NCT06964412).Results:
Between July 2025 and February 2026, 36 patients and 16 new learners were enrolled (13 residents, 3 fellows/new attendings). The mean DSC between learner and attending contours was 0.84 (standard deviation (SD) = 0.09) with manual contouring and 0.93 (SD = 0.06) with AI assistance (p = 0.0005). Time from scan acquisition to final contour approval was numerically shorter with AI (19.7 vs 23.3 minutes, p = 0.22).Conclusion:
To our knowledge, this is the first prospective, randomized, blinded evaluation of AI-assisted contouring within a trainee educational workflow. AI-assisted learner contours were more similar to attending contours, meeting our predefined primary endpoint. This suggests AI improves contouring accuracy among new learners, with potential implications for training efficiency, reduced learning curve, and broader adoption of complex radiation techniques.