Machine Learning-Based Support for Monitor Unit and Lung Shielding Estimation in Conventional Total Body Irradiation.
Authors
Affiliations (3)
Affiliations (3)
- Department of Oncology, University of Turin, 10126 Turin, Italy.
- Medical Physics Unit, A.O.U. Città della Salute e della Scienza di Torino, 10126 Turin, Italy.
- Radiation Oncology Unit, Department of Oncology, A.O.U. Città della Salute e della Scienza di Torino, 10126 Turin, Italy.
Abstract
<b>Background/Objectives</b>: Total body irradiation (TBI) is widely used in conditioning regimens before hematopoietic stem cell transplantation. In conventional opposed-field TBI, monitor unit (MU) calculation and lung shielding definition are often based on manual procedures that may introduce operator-dependent variability. This study aimed to develop machine learning (ML) models to support the prediction of these treatment parameters using routinely available clinical and imaging data. <b>Methods</b>: A retrospective analysis was performed on 80 patients treated with conventional opposed-field TBI. Clinical, geometric, and CT-derived variables were used to train regression models for MU prediction. Feature selection was performed using LASSO regression, followed by Ridge regression for final modeling. Lung shielding thickness prediction was developed using planning CT data from 66 patients through recursive feature elimination and Random Forest regression. Model performance was assessed using nested 5-fold cross-validation and mean absolute error (MAE). <b>Results</b>: The final Ridge model achieved an MAE of 74.0 ± 6.9 MU, improving compared with the full-feature model 115.6 ± 44.0 MU. The Random Forest benchmark achieved an MAE of 81.1 ± 10.3 MU. For lung shielding thickness prediction (6-9 mm), the Random Forest model achieved an MAE of 0.60 mm. Prediction uncertainties were consistent with clinically accepted in vivo dosimetric tolerances. <b>Conclusions</b>: ML-based models can support the estimation of key TBI treatment parameters, potentially improving workflow efficiency and reducing operator-dependent variability while complementing standard treatment planning and verification procedures.