49 - Automated Prostate Segmentation for MRI-guided Transurethral Ultrasound Ablation: Validation and Real-world Experience

Sunday, April 12, 2026

4:12 PM - 4:21 PM ET

Presenting Author(s)

Steven S. Raman, MD, FSIR

Professor of Radiology, Urology and Surgery
David Geffen School of Medicine At UCLA, United States

Author/Co-author(s)

LK

Laurence Klotz, MD, FRSC

Professor of Surgery
Sunnybrook Health Sciences Centre, Canada
SA

Sandeep Arora, MD

Associate Professor of Radiology and Biomedical Imaging
Yale School of Medicine, United States
Pejman Ghanouni, MD, PhD

Associate Professor of Radiology
Stanford University Medical Center, United States
AM

Adnaan Moin, MD

Interventional Radiologist
WellSpan Health, United States
PS

Preston Sprenkle, MD

Associate Professor of Urology
Yale School of Medicine, United States
CP

Christian Pavlovich, MD

Professor of Urology
Johns Hopkins University School of Medicine, United States
AK

Ara Karamanian, MD

Vascular and interventional radiologist
Halo Diagnostics, United States
AO

Aytekin Oto, MD, MBA

Professor of Radiology and Surgery
University of Chicago, United States
ES

Elyas Shaswari, PhD

Data Scientist
Profound Medical, Canada
BL

Ben Leung, MS

Engineering and Clinical Scientist
Profound Medical, Canada
CW

Cameron Wright, PhD

Data Scientist
Profound Medical, Canada
RP

Robert Princenthal, MD

Radiologist and Director of RadNet Prostate Program
Radnet Imaging, United States
ES

Edward Steiner, MD, FACR

Chief, Advanced Prostate Care Center
WellSpan Health, United States

Purpose: MRI-guided transurethral ultrasound ablation of the prostate (TULSA) uses intraoperative imaging to guide ultrasound-induced thermal coagulation of a physician-defined target volume. Segmentation of the ablation zone across multiple prostate sections is repetitive, time-consuming, and subject to reader variability. We describe validation of a deep learning whole-gland segmentation model integrated into TULSA treatment planning software, performance improvements with an updated model, and the real-world experience with AI-assisted contouring on planning time.

Materials and Methods: Two AI prostate segmentation models (first using SEResNet34, later SAM2-UNet) were trained on T2w MRI from public datasets and past TULSA cases. The first model was validated in a multi-reader multi-case study; the second was benchmarked against those results. In the validation study, 3 urologists and 5 radiologists segmented images from 51 past TULSA cases, once manually and once with AI-assistance, separated by a 4-week washout. Dice similarity scores were calculated against a reference segmentation defined using the STAPLE method by 3 other expert radiologists. Segmentation times were also compared. After regulatory clearance, treatment planning times were measured in 153 real-world cases with AI versus 962 past cases without AI at the same institutions. Physicians were surveyed on segmentation quality.

Results: Segmentation accuracy of the original model was 0.914±0.027. Among urologists, accuracy and speed of prostate segmentation improved with AI assistance vs. manual contouring (Dice 0.919±0.020 vs. 0.909±0.021, and time 3.6±2.5 vs. 5.3±3.3 min, both p< 0.0005). Radiologists had equivalent accuracy but were 30% faster with AI vs. without (0.927±0.021 vs. 0.928±0.020, and 2.5±1.4 vs. 3.6±1.4 min). The later SAM2-UNet model had standalone accuracy of 0.921±0.023, superior to urologists (p< 0.0005) and non-inferior to radiologists. In real-world use, AI reduced planning time by 11% (p=0.005). Physicians reported for treatment-naïve prostates, AI accuracy was sufficient to reduce time and effort required to define a treatment plan.

Conclusion: AI prostate segmentation in TULSA planning resulted in non-inferior segmentation accuracy and significantly reduced planning time. The updated model had non-inferior accuracy to radiologists and was superior to urologists. Real-world results support the time savings with AI-assisted planning.