Iterative Refinement of Radiation Therapy Dose Distribution Prediction for Accelerated Partial Breast Irradiation via Plan Scoring

The integration of artificial intelligence into clinical workflows offers substantial opportunities to enhance treatment quality in radiation oncology but also require methods that address the complexity of clinical decision making. Recent advances in dose prediction demonstrate the value of conditioning on patient specific anatomy. However, the tendency of deep learning models to regress toward cohort averages motivates strategies that preferentially learn from higher quality exemplars. Building on these insights, we present a framework that uses clinically relevant metrics to refine prediction quality for accelerated partial breast irradiation. A linear piecewise scoring system assigns normalized scores to each plan across dose volume histogram metrics covering target coverage and sparing of organs at risk. We applied this framework to a retrospective cohort of 550 patients treated at our institution, and we trained a three dimensional dose prediction model based on a hierarchically dense U Net. Preliminary results reveal wide variation in score distributions across the cohort, reflecting both patient specific complexity and variation in planning quality. We trained a baseline model and then applied our iterative refinement framework, which ranks cases by the composite quality score and, at each round, retains the highest scoring half for retraining. Models refined in this manner demonstrated promising improvements in predicted plan quality, supporting the effectiveness of this approach. These findings illustrate how scoring guided curation can align model behavior with clinical priorities and provide a path toward adaptive, high precision treatment planning tools.