Evaluation of AI and radiologist contouring in prostate MRI for targeted MRI/US fusion biopsy

Abstract

Objective: Prostate cancer is an increasingly prevalent public health issue, particularly in aging populations such as Thailand. While traditional diagnostic methods like systematic transrectal ultrasound-guided biopsy are widely used, they can result in overdiagnosis and unnecessary treatment. MRI/Ultrasound (MRI/US) Fusion Biopsy offers greater precision by targeting suspicious areas detected in MRI scans. However, manual contouring of the prostate and lesion locations by radiologists or urologists is time-consuming and subject to variability, potentially delaying diagnosis and treatment.

Materials and Methods: This retrospective study developed and evaluated an AI-based prostate segmentation model using 125 annotated prostate MRI cases (3,193 images) from a public dataset for training, and then it was tested on 109 clinical cases (2,952 images) from the National Cancer Institute. The model combined a YOLO-based bounding box detection with the segment anything model (SAM) for prostate segmentation. Model performance was compared to radiologist-drawn contours using dice similarity coefficient (DSC) and % relative percent difference (RPD) in prostate volume estimation.

Results: For cases not requiring post-processing, the AI model achieved a mean DSC of 0.72 and an RPD of 8.90% in comparison to radiologist contours. For cases requiring post-processing, the DSC dropped to 0.66 and the RPD increased to 13.45%. These results indicate a high level of agreement between the AI and expert annotations, particularly in standard cases.

Conclusion: The AI-based model demonstrated promising accuracy with regard to segmentation of the prostate gland on MRI scans, comparable to radiologist performance. This approach has the potential to reduce diagnostic delays and lessen the workload of radiologists in prostate cancer workflows. Future improvements should focus on enhancing model precision, incorporating prostate imaging-reporting and data system (PI-RADS) scoring, and validating the system across diverse clinical settings to support safe and effective integration into routine diagnostic practice.