2648 - Evaluation of Machine Learning-based Automated Proton Therapy Planning in Locally Advanced Breast Cancer

Purpose/Objective(s): To evaluate the feasibility and dosimetric accuracy of machine learning (ML) generated treatment plans for patients with locally advanced breast cancer undergoing pencil beam scanning proton radiotherapy.

Materials/Methods: A total of 140 proton therapy treatment plans for patients with locally advanced breast cancer, previously treated at our institution using the MEVION S250i system, were retrospectively collected to establish a plan database. Of these, 111 plans were allocated for training a three-dimensional dense U-Net deep learning model designed to predict three-dimensional dose distributions. Model inputs comprised the prescribed dose, target volume contours (clinical target volume [CTV] of the chest wall, internal mammary lymph nodes, axillary lymph nodes, and supraclavicular lymph nodes), the external body contour, and four organs-at-risk (OAR) contours: spinal cord, heart, lungs, and esophagus. The predicted dose distribution was subsequently used as a reference for voxel-wise dose mimicking and inverse optimization within the RayStation treatment planning system, facilitating automated generation of a clinically deliverable treatment plan. Plan quality was evaluated by comparing dose volume histogram (DVH) metrics between the machine learning generated plans and the corresponding clinical plans in an independent cohort of 29 breast cancer cases.

Results: ML-generated plans demonstrated dosimetric quality comparable to clinical plans. Mean absolute dose differences to the CTV chest wall were 225 ± 40cGy, with D99% and D95% differences of -46 ± 35cGy and 54 ± 18cGy, respectively. For organs at risk, mean ± standard deviation differences were: heart Dmean 50 ± 6cGy, Dmax -568 ± 61cGy; lung Dmean 91 ± 9cGy, Dmax -11 ± 34cGy; contralateral chest wall Dmean 16 ± 26cGy, and Dmax -56 ± 83cGy.

Conclusion: Machine learning generated treatment plans for locally advanced breast cancer using pencil beam scanning proton therapy are feasible and produce dosimetric quality comparable to clinically delivered plans. The 3D deep learning model accurately predicted dose distributions that, when used for automated plan generation, maintained target coverage and organ-at-risk sparing within clinically acceptable limits. This approach demonstrates potential to streamline treatment planning, support individualized plan evaluation, and improve planning consistency and efficiency.