Evaluation of a Third Party AI Auto-Contouring Tool for Online MRI Guided Adaptive Radiotherapy In Prostate SBRT
Abstract
Purpose
MRI guided Adaptive Radiation Therapy (MRgART) allows for daily plan adaptation to the patient’s anatomy. This is achieved by imaging the patient before each fraction, editing the contours to the new anatomy, and adapting the plan to the new contours. However, since the patient is on the table, time is a concern because of potential patient discomfort and drifts in internal anatomy after the initial image is taken. To limit treatment time, online contour edits are often limited to specific structures or contours within a certain proximity of the target. Therefore, the plan is adapted to structures that do not always fully represent the true anatomy. This project evaluates the utility of a third party AI auto-contouring tool to assist clinicians in the online workflow, allowing for more accurate contouring without increasing treatment times.
Methods
Twenty anonymized MRgART prostate SBRT patients’ daily images, structures, and treatment plans were gathered and anonymized for this study. The daily images were processed by the auto-contouring tool, which contoured the organs at risk (OARs). The autogenerated OAR contours were compared to the corresponding clinically used contours, and dosimetric impact was evaluated by recalculating the initial plan on the new contours.
Results
The largest and smallest average percentage volume difference between the auto-contoured and clinical structures was the penile bulb and rectum respectively. All mean DICE scores were <0.800, except the external and bone contours. Recalculated plans had at least one failed scorecard constraint in 52% of fractions.
Conclusion
Discrepancies were found between the clinical and auto-contoured OAR structures, reflecting limited editing during the online workflow. Dosimetric results indicate plan quality may be improved when utilizing the auto-contours. We hypothesize this auto-contouring tool produces more accurate OAR contours compared to the current online workflow, supporting more accurate plan generation, without increasing treatment times.