3736 - A Deep Learning-Driven Automated Radiotherapy Treatment Planning Pipeline for Breast Cancer: Feasibility and Clinical Evaluation

Purpose/Objective(s): This study aimed to evaluate the clinical feasibility of an automated radiotherapy treatment planning pipeline for breast cancer, focusing on optimizing planning efficiency, minimizing inter-planner variability, and maintaining or enhancing treatment plan quality compared to traditional manual planning methods.

Materials/Methods: The developed pipeline integrates a 3D U-Net-based dose prediction model and a multi-objective optimization framework. Separate models were trained for left- and right-sided breast cancer cases (50 cases per side, 47.5 Gy prescription). The models utilized contoured targets (PTVs) and organs-at-risk (OARs) to predict 3D dose distributions. Given the anatomical differences between left and right-sided cases, distinct training models were employed for each. The optimization phase incorporated predicted doses as constraints, with three strategies: (1) individualized target handling with auxiliary rings to improve dose gradients, (2) a priority hierarchy to resolve conflicts, emphasizing critical OAR constraints (e.g., esophagus Dmax), and (3) flexible modulation of predicted doses to refine constraint strength. The pipeline was tested on 30 independent cases (15 cases per side) using a commercial treatment planning platform.

Results: The dose prediction models achieved mean absolute errors of 1.40 Gy (left) and 1.50 Gy (right). Automated plans demonstrated comparable or superior performance in target coverage and heart sparing. Notably, doses to the contralateral breast and esophagus were significantly reduced due to the auxiliary rings and priority mechanisms. While V_5Gy of the ipsilateral lung increased slightly for right-sided cases (about 3%), it remained clinically acceptable. Additionally, the planning time decreased from 48.2±13 minutes to just 5.0±0 minutes.

Conclusion: This fully automated planning pipeline demonstrates clinical feasibility and significant improvements in planning efficiency without compromising the plan quality. It shows potential for standardizing planning processes and reducing dosimetrist workload, particularly in resource-limited settings. Future work will focus on validating the pipeline's generalizability across multiple disease types and institutions.