SIR Today
New AI model offers fast, accurate prostate segmentation
Presentation: Sunday, April 12, at 3 p.m. during the Men’s Health 1 session
S. Raman, L. Klotz, S. Arora, P. Ghanouni, A. Moin, P. Sprenkle, C. Pavlovich, A. Karamanian, A. Oto, E. Shaswari, B. Leung, C. Wright, R. Princenthal, E. Steiner
A new AI model for prostate segmentation during transurethral ultrasound ablation (TULSA) may offer an accurate and time-saving alternative to manual segmentation, according to researchers.
The TULSA procedure provides a treatment option for men with prostate cancer and/or benign prostatic hyperplasia that emphasizes precision and preservation. A key step of this process is prostate segmentation, which can introduce variability and delays.
“Prostate segmentation is a critical but repetitive and time-intensive step in treatment planning for the TULSA procedure,” said Ben Leung, an engineering and clinical scientist with Profound Medical, who worked on the AI model. “As an MRI-guided procedure that generates rich imaging data, TULSA is well suited for AI-enabled support. We saw an opportunity to leverage AI to improve workflow efficiency while enhancing consistency and predictability in treatment planning.”
According to Leung, the integration of AI has the potential to further standardize the procedure, improve workflow efficiency and support more consistent clinical outcomes—reinforcing the role of advanced imaging and intelligent tools in the future of prostate care.
Researchers examined the impact of AI-assisted tools on the behavior and performance of radiologists and urologists during prostate segmentation, specifically to assess whether AI-assistance could meaningfully reduce the time required to complete segmentation while maintaining non-inferior accuracy compared to manual methods.
The researchers trained two separate models on thousands of prostate MRIs—both public datasets and real‑world TULSA cases. In a controlled study, three urologists and five radiologists each contoured 51 past TULSA scans twice—once by hand and once with the AI‑generated outline that they could edit. A four‑week washout period kept the two sessions independent.
After regulatory clearance, researchers say, treatment planning times were measured in 153 real-world cases with AI versus 962 past cases without AI at the same institutions.
A review of the data showed that the AI model saved urologists approximately 5.3 minutes in contouring time, while radiologists reduced their time by nearly 30%, from 3.6 minutes to 2.5 minutes. Accuracy did not meaningfully improve for radiologists using the model, though urologists improved their Dice scores from 0.909 to 0.919.
“Our findings suggest that integrating AI into the TULSA procedure workflow enables more consistent treatment planning while reducing overall procedure and anesthesia times,” Leung said. “This has meaningful benefits for both physicians and patients, supporting efficiency without compromising accuracy or clinical confidence.”
According to Leung, future research will focus on expanding AI-assisted capabilities, such as prostate zonal segmentation, segmentation of surrounding anatomy, and lesion detection.
“These advancements could further support physicians in customizing treatment plans and optimizing outcomes for individual patients,” he said.
Related
![]()
Related
![]()
Related