Clinical implementation of deep learning contour autosegmentation for prostate radiotherapy

Area of Science:

• Radiotherapy
• Medical Imaging
• Artificial Intelligence

Background:

• Clinical evaluations of artificial intelligence (AI)-driven autosegmentation in radiotherapy are limited.
• This study addresses the need to assess the clinical utility of deep learning autosegmentation for MRI-based prostate cancer radiotherapy planning.

Purpose of the Study:

• To evaluate the clinical utility and efficiency of deep learning-based autosegmentation for Magnetic Resonance (MR)-based prostate radiotherapy planning.
• To compare automated contours with physician-drawn contours using geometric indices and physician feedback.

Main Methods:

• Prospective data collection from prostate-only radiation patients (June-December 2019).
• Geometric indices (VDSC, SDSC, APL) compared automated and final contours.
• Physician contouring time and protocol deviation ratings were recorded.

Main Results:

• 173 patients included; 85% received SBRT. CTV (prostate and seminal vesicles) showed median VDSC of 0.89 and SDSC of 0.91.
• Physician surveys (78% response rate) indicated 33% of autocontours required major edits.
• Autosegmentation resulted in a 30% time saving (12 minutes) in physician contouring time.

Conclusions:

• Deep learning autosegmentation is effective for prostate radiotherapy planning, enhancing efficiency.
• Major edits were uncommon and geometric indices showed weak correlation with contouring time or quality.
• Added Path Length (APL) is a clinically relevant metric; further work needed on physician education and quality assurance.