Convex Imputing for Automatic Treatment Planning
G Sayre*, D Low, D Ruan, UCLA, Los Angeles, CASU-E-T-639 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
Purpose: To improve plan quality and consistency and to automate the treatment planning process by developing a novel data-mining method that learns a task-specific objective formulation from previously optimized cases using a convex imputing approach.
Methods: General clinical planning objectives are hard to quantify and DVH conditions are challenging to optimize directly, often requiring numerous iterations of human intervention that incurs manpower cost and risk of practice inconsistency. In this work, we address treatment planning using a new data-mining perspective that learns the weighting factors of user-defined convex penalties from previously optimized cases. The imputed weighting factors and corresponding penalties are then used to optimize future cases. To test our method, we imputed weighting factors from clinically-derived prostate cases. The imputed weighting factors were then inserted into an optimization engine to drive generation of new plans. DVHs were then calculated from the imputed plans and compared to DVHs of the clinical plans.
Results: Dose-volume metrics calculated from the imputed plans showed excellent correlation with the original plans for each structure of interest: 1) PTV coverage R^2 = 0.782, 2) V_R_70Gy R^2 = 0.968, 3) V_R_65Gy R^2 = 0.983, 4) V_R_60Gy R^2 = 0.989, 5) V_R_50Gy R^2 = 0.9952, 6) V_Bl_75Gy R^2 = 0.857, 7) V_Bl_70Gy R^2 = 0.9989, 8) V_Bl_65Gy R^2 = 0.9993, and 9) bulb mean dose R^2 = 0.9776. Furthermore, the inter-case averages of these metrics deviated by at most 3 units and on average by 0.84 units. All imputed plans were clinically acceptable.
Conclusion: This pilot study demonstrates the feasibility and great potential of utilizing the developed convex imputing approach to achieve dose plans that match the true planning goals of a dose planner and/or an institution.
Funding Support, Disclosures, and Conflict of Interest: This work is supported in part by an AACR career development award and a TRDRP exploratory grant.