Artificial intelligence is increasingly utilized to generate contours for organs at risk and even target volumes. These deep learning autosegmentation (DLAS) models are often trained using a vendor or institution’s own dataset. But for more widespread utilization, the question is whether these models can be retrained to better match the practices of an individual institution? In this study, a vendor DLAS model for contouring pelvic OARs and the prostate fossa CTV was retrained using institution specific cases. Only 30-60 institutional cases were required to generate high quality, acceptable OAR contours. This was particularly true of the femoral heads and bladder with slightly more difficulty in generating consistent high quality contours for the penile bulb and rectum. The CTV was more difficult, though. Even though the DLAS model performed better with an increasing number of institutional training cases, the CTVs were graded as precise or acceptable by physician reviewers for just 54% of DLAS delineations compared to 73% of manual delineations. | Hobbis, Pract Radiat Oncol 2023
