Anatomical Parameter-driven Volumetric Modulated Arc Therapy Optimization in Left-sided Breast Cancer: A Machine Learning Framework for Lung Dose Prediction

Purpose: The aim of this research is to assess different volumetric modulated arc therapy (VMAT) methods employed in the radiotherapy treatment of left-sided breast cancer, as well as to develop a predictive model for lung doses by leveraging machine learning techniques. Materials and Methods: In this retrospective study involving 101 patients with left-sided breast cancer, we conducted a comparison of three VMAT techniques: two-partial arc (VMAT-2P), VMAT-4P, and VMAT-5P. We assessed various anatomical parameters, including the central lung distance (CLD), the tangent left lung volume (LLV) within the treatment area, the LLV, and the planning target volume (PTV). In addition, we established predictive models for lung doses utilizing the Least Absolute Shrinkage and Selection Operator, ridge, and linear regression algorithms while also evaluating dosimetric parameters for both the PTV and the organs at risk (OARs). Results: The VMAT-5P technique resulted in the lowest exposure to low doses in the lungs (V5 Gy: 56.50%–58.00%), whereas VMAT-4P exhibited the lowest mean lung dose (11.83–12.65 Gy) and V20 Gy values (21.79%–25.07%). CLD was identified as the most critical predictor of lung dose (relative importance: 0.45–0.60). The Random Forest Regressor outperformed other models in predicting mean lung dose ( R 2: 0.235–0.341), whereas regularized linear models showed the greatest resilience to variations in input data. Conclusion: VMAT-4P is identified as the most effective method for radiotherapy in left-sided breast cancer, providing an excellent balance between optimal target coverage and improved protection of surrounding OAR. The Random Forest Regressor provided the most accurate mean lung dose predictions.