GPU-Based Automatic Treatment Planning Using Previously Delivered Treatment Plans as Prior Knowledge
N Li1,2*, Q Gautier1, M Zarepisheh1, Y Graves1, Z Tian1, L Zhou2, X Jia1, K Moore1, S Jiang1, (1) Center For Advanced Radiotherapy Technologies,University of California, San Diego,CA, (2) Southern Medical University, Guangzhou, GuangdongSU-E-T-652 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
To demonstrate the feasibility of automatically and efficiently developing a clinically optimal treatment plan for a new patient using a GPU-based automatic planning engine and a library of previously delivered treatment plans.
An in-house GPU-based platform called SCORE was used for automatic treatment planning. A library of 65 prostate IMRT plans previously delivered at our institution has been assembled. Leave-one-out cross validation was performed whereby each of the 65 patients was selected as a 'new' patient and the other 64 patients were treated as reference patients. A treatment plan was automatically generated for the 'new' patient using the same beam setup and guided by the DVH of the reference plan. A set of candidates were generated by filtering the 64 plans using institutional criteria for treatment plan quality. A GUI was developed to allow clinicians to navigate through the candidate plans at high efficiency to select the best plan.
For each prostate IMRT patient, it took 40-60 minutes to generate 64 new plans without any manual intervention. Compared with the original clinical plan created through the regular planning process, each of the final plans chosen from the candidates had equal or better plan quality in terms of DVH curves and specific plan quality metrics (D95% and Dmax for PTV, V65 and V40 for rectum and bladder, Dmax for femoral heads).
We proposed a GPU-based automatic treatment planning procedure as a solution to the current manual treatment planning process requiring significant human effort, planner experience, and clinician. By using this procedure that leverages prior planning experience, a set of candidate plans can be generated without any manual intervention from which it is easy for the clinician to find a clinically optimal plan using an interactive plan selection GUI.