Assessing AI Contouring Efficiency in Male Pelvic MRI and CT Imaging

open_access
Open Access
Author:
Maheshwari, Aditi
Area of Honors:
Physics
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Kohta Murase, Thesis Supervisor
Anna Stasto, Thesis Honors Advisor
Keywords:
AI Contouring
Radiology
Radiotherapy
Artificial Intelligence
Auto-Contour
MRI
Magnetic Resonance Imaging
Abstract:
Purpose: To evaluate the clinical accuracy and efficiency of artificial intelligence (AI)–based auto-contouring for delineating and identifying male pelvic organs at risk on both computed tomography (CT) and magnetic resonance imaging (MRI) in radiotherapy planning, and to assess the potential role of MRI-based AI segmentation in supporting the ongoing transition toward MRI- guided radiotherapy workflows. Methods: A commercially available deep learning auto-contouring system (AI-Rad Companion Organs RT, Siemens Healthineers) was applied to paired CT and MRI datasets from 20 male patients at Penn State College of Medicine. Structures of interest included the prostate, bladder, rectum, seminal vesicles, and penile bulb. Two board-certified genitourinary radiation oncologists independently reviewed all AI-generated contours using a standardized four-point qualitative scoring system and recorded contour review and editing times. Contours receiving scores of 3 (minor edits required) or 4 (clinically acceptable without modification) were considered clinically positive. Interobserver agreement was assessed on cases reviewed by both physicians. AI-generated prostate volumes on CT and MRI were extracted and compared to characterize modality-specific volumetric differences. Results: Qualitative scoring was completed across 16 CT datasets (Physician 1 n = 4, Physician 2 n = 12) and 14 MRI datasets (Physician 1 n = 4, Physician 2 n = 10). Overall clinical acceptability was 72.5% on CT and 82.9% on MRI. On CT, the bladder and penile bulb achieved the strongest performance (100% and 93.8% acceptability, respectively), while the rectum (31.2%) and prostate (56.2%) were the lowest-performing structures. On MRI, the penile bulb and bladder again led performance (92.9% each), with the rectum improving markedly to 78.6% compared with CT. The prostate showed the greatest interphysician variability, with Physician 1 rating 0% of prostate contours as acceptable on both modalities and Physician 2 rating 75.0% and 90.0% acceptable on CT and MRI, respectively. Mean contour review times were 30.5 minutes (CT) and 21.2 minutes (MRI) for Physician 1, and 10.5 minutes (CT) and 4.3 minutes (MRI) for Physician 2. Interobserver agreement was highest for the rectum (mean absolute score difference 0.00) and lowest for the prostate and penile bulb (0.75). AI-generated prostate volumes were larger on CT than MRI in 84.2% of cases, with a mean CT overestimation of 27.0% relative to MRI. Conclusion: AI-based auto-contouring demonstrated clinically acceptable performance for the majority of male pelvic structures on both CT and MRI, with the prostate and rectum representing consistent areas requiring physician intervention. The substantial interphysician variability in both contour scoring and review time underscores the importance of structured review workflows. AI-generated CT prostate volumes systematically exceeded MRI volumes, with direct implications for treatment planning accuracy. These findings support the role of AI-assisted contouring as a starting-point reference in male pelvic radiotherapy planning and demonstrate its feasibility in MRI-based workflows as the field continues to integrate MRI into treatment planning and MRI-only workflows [1].