Machine learning prediction for lung dose in locally advanced esophageal cancer using volumetric modulated arc therapy

Abstract

We developed machine learning (ML) models for predicting lung dose-volume histogram (DVH) metrics [V_5 Gy, V_20 Gy, and mean lung dose (MLD)] in locally advanced esophageal cancer volumetric modulated arc therapy and assessed the prediction accuracy of the models. Four ML models (linear regression, support vector machine, decision tree, and ensemble) were built with fivefold cross-validation of the predicted lung DVH metrics using a developed program by MATLAB R2022a. Eight explanatory variables were employed: gender, with/without simultaneous integrated boost and jaw tracking, age, height, weight, the ratio of the total irradiation angle to the total rotation angle of the gantry, and the ratio of the longitudinal length of the planning target volume overlapped with the whole lung to the length of the whole lung. To evaluate the prediction accuracy of the ML models, the differences and the Pearson correlation coefficients (r) between the predicted and planned doses were calculated. The mean ± standard deviation values of the planned lung doses of V_5 Gy, V_20 Gy, and MLD were 34.9 ± 15.2%, 11.9 ± 6.7%, and 7.2 ± 3.3 Gy, respectively. The differences for all models were −0.1 ± 8.0% (V_5 Gy,), 0.1 ± 4.2% (V_20 Gy), and −0.2 ± 1.7 Gy (MLD). The predicted lung doses were consistent with the clinically planned doses (V_5 Gy [r = 0.7–0.8], V_20 Gy [r = 0.6–0.8], and MLD [r = 0.7–0.9]), and there was no significant difference in the prediction accuracy among the ML models. These models can promptly evaluate and improve the quality of treatment plans by aiding patient-specific decision-making regarding lung-dose reduction before treatment planning.