In a phase II trial, AI-assisted contouring for prostate cancer brachytherapy improved contour quality among residents, fellows, and junior attending radiation oncologists, producing contours that more closely matched those approved by experienced attending physicians.  

“AI-assisted learner contours were significantly more concordant with attending-approved contours and have met our predefined primary endpoint. We also saw improvement in workflow time, suggesting AI may improve both contouring accuracy and procedural efficiency,” said Bryce Thomsen, MD, of Brigham and Women’s Hospital and Dana-Farber Cancer Institute, who presented findings from the randomized, blinded study at the 2026 American Society for Radiation Oncology (ASTRO) Annual Meeting (Abstract 1074).

“These findings suggest that AI can improve contour quality, reduce workflow burden, and potentially support broader adoption of high-dose-rate prostate brachytherapy,” Dr. Thomsen added. 

Background

Brachytherapy is an effective treatment option for patients with prostate cancer. However, its use has declined, and trainees have reported low confidence in performing high-dose-rate brachytherapy.  

According to a survey of graduating radiation oncology residents about their confidence in performing various radiation techniques and modalities, only one in four felt confident enough to perform high-dose-rate prostate brachytherapy upon graduation. Confidence levels were similar for low-dose-rate brachytherapy. Low- and high-dose-rate brachytherapy ranked third and fourth lowest, respectively, in terms of resident confidence.  

Dr. Thomsen suggested that this lack of confidence may be attributable to the “deep learning curve associated with TRUS [transrectal ultrasound]–based prostate contouring.” He presented examples of trainee-generated contours illustrating difficulties in distinguishing needle reverberation artifacts and gland swelling.  

The researchers hypothesized that AI could improve contouring quality and procedural efficiency among novice practitioners performing prostate brachytherapy.  

Study Methods

“This is, to our knowledge, the first prospective randomized blinded trial evaluating AI-assisted contouring embedded directly within a live brachytherapy workflow,” Dr. Thomsen noted.  

The researchers conducted a randomized, blinded phase II trial at a single institution. Eligible patients were candidates for high-dose-rate prostate brachytherapy. Participating learners were residents, fellows, or junior attending radiation oncologists who had completed fewer than 20 TRUS cases.  

The new learners were randomly assigned 1:1 to manual contouring or AI-assisted contouring, in which an algorithm automatically segmented the prostate and organs at risk before the learners reviewed and edited the contours. In both arms, a blinded attending radiation oncologist reviewed and adjusted the contours before treatment delivery. Learner-generated contours were then compared with attending-approved contours using the prostate Dice similarity coefficient on TRUS images. Secondary endpoints included workflow efficiency, expert-rated contour quality, and learner-reported experiences with the AI system.  

A total of 36 patients underwent prostate brachytherapy, with manual contouring performed in 18 cases and AI-assisted contouring in 18 cases. Most patients (61%) received high-dose-rate brachytherapy boost or salvage monotherapy. 

The participating learners included 13 residents and 3 fellows or junior attending radiation oncologists. Among residents, 85% had previously contoured five or fewer cases, and 75% of all learners reported feeling uncomfortable with TRUS-based contouring at study entry. Each learner contoured a median of two cases.  

Study Findings

The study met its primary endpoint, with a mean Dice similarity coefficient of 0.93 in the AI-assisted contouring arm vs 0.84 in the manual contouring arm (P = .0005). Dr. Thomsen noted that this difference was statistically significant and “demonstrates that AI improved contour accuracy among the new learners.” 

He explained that the difference between the treatment arms was driven primarily by residents, who had less brachytherapy experience than the other learners. Among residents, the mean Dice similarity coefficient was lower than the overall mean in the manual contouring arm. Among fellows and junior attending radiation oncologists, the mean coefficient was higher than the overall mean in the AI-assisted arm.  

Mean contouring time was 18.5 minutes (standard deviation = 7.7) in the AI-assisted arm vs 24.2 minutes (standard deviation = 8.5) in the manual arm, representing a 24% reduction (P = .04).  

In the post-contouring survey, most learners reported that AI was educationally valuable. Dr. Thomsen suggested that AI could serve as a scaffold rather than a shortcut in trainee education.  

DISCLOSURES: The study authors reported no conflicts of interest. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.