Abstract
Esophageal cancer (EC) is among the most common malignancies worldwide, with radiotherapy as a key treatment. Intensity-modulated radiation therapy (IMRT) offers highly conformal dose distributions while sparing critical organs; however, planning is complex due to the esophagus' central location and proximity to vital structures. This study aimed to develop a personalized, automated IMRT planning system for EC. A retrospective analysis of 301 patients treated with 45 Gy in 25 fractions was conducted. Patient data, including CT images, structure sets, and clinical plan (CP) dose distributions, were collected to develop and internally validate the model. Geometric predictors of lung and heart doses were identified based on planning target volume (PTV) shape, size, and proximity to organs. External validation was subsequently performed on an independent cohort of 20 patients. Predictive models were built using univariate linear regression. The auto-planning (AP) system, developed via Python interfaced with the Eclipse Scripting Application Programming Interface (PyESAPI). Among all geometric predictors, LungCropPTV(V)/Lung(V) and HeartCropPTV(V)/Heart(V) were significantly correlated with lung and heart dosimetric endpoints, with R² of 0.771, and 0.934, respectively (both p < 0.001). In internal validation, AP generated plans in approximately 4 min, reduced lung V(20) by 13-21.3%, and lowered monitor units (MUs) compared to CP. External validation confirmed comparable performance. This personalized AP demonstrated performance in reducing lung dose and enhancing treatment precision, with potential to improve EC radiotherapy outcomes pending further clinical validation.