3203 - Deep Learning-Based Pre-Assessment for Thoracic Radiotherapy Auto-Planning: Predicting Plan Achievability before Planning
Presenter(s)
H. Zhai1, J. Wang2, and W. Hu2; 1Fudan University Shanghai Cancer Center, Shanghai, China, 2Department of Radiation Oncology, Fudan University Shanghai Cancer Center; Department of Oncology, Shanghai Medical College, Fudan University; Shanghai Clinical Research Center for Radiation Oncology; Shanghai Key Laboratory of Radiation Oncology, Shanghai, China
Purpose/Objective(s): This study aimed to develop and validate an AI model that predicts the achievability-related score (ARS) of radiotherapy plans. It incorporates target geometric characteristics, prescription dose, and the subjective difficulty scores from experienced medical physicists. The model aims to optimize the radiotherapy workflow and enhance resource allocation in treating thoracic tumors by automating the assessment of plan difficulty.
Materials/Methods: This study presents a retrospective analysis involving 400 patients diagnosed with thoracic tumors who received radiotherapy at Fudan University Shanghai Cancer Center during the period from November 2024 to April 2025. Five medical physicists assigned subjective difficulty scores (1–5) based on the spatial relationships between the PTV and OARs as well as the prescription dose. These expert-assigned scores served as the ground truth labels. The PTV, five OARs, and prescription dose information were processed as seven-channel inputs to train and evaluate a 3D ResNet model. Using five-fold cross-validation, the model was developed to predict the ARS of radiotherapy plans. To enhance model interpretability, 18 geometric–dosimetric indices were extracted to identify key factors influencing predicted complexity. Furthermore, the model was prospectively validated by comparing the clinical goal attainment rates of 20 extreme cases (10 highest vs. 10 lowest ARS) generated via a standardized auto-planning template.
Results: Across 400 thoracic radiotherapy plans, five physicists showed excellent scoring consistency (ICC = 0.856). Five-fold cross-validation demonstrated stable model performance: mean absolute error (MAE) = 0.377, root mean square error (RMSE) = 0.464, coefficient of determination (R²) = 0.716, and correlation coefficient (Pearson r = 0.857, Spearman ? = 0.843). Spearman correlation analysis identified 13 significant geometric-dosimetric factors (p < 0.05), with PTV volume and the overlap volume between PTV and OARs acting as the strongest positive drivers, while Sphericity Index (SI) served as a key negative driver of plan difficulty. In the prospective cohort (n = 20), high-ARS plans exhibited significantly poorer dose homogeneity than low-ARS plans (p < 0.05), whereas conformity remained comparable between groups (p = 0.345). Nevertheless, one low-ARS case failed to meet all institutional clinical criteria, while zero cases in the high-ARS group achieved acceptability.
Conclusion: This study establishes a deep learning-based model that effectively predicts radiotherapy plan achievability prior to the planning process. By providing high-accuracy difficulty assessments, this model serves as a reliable a priori triage tool to optimize clinical workflows and improve resource allocation. Specifically, it enables the proactive identification of complex cases requiring expert intervention, thereby bridging the gap between automated planning and individualized treatment strategies.