Predictive Modeling of Computed Tomography Facility Radiation Shielding Requirements Using Machine Learning Regression and Barrier Risk Classification

Abstract

Ionizing radiation from computed tomography (CT) examinations poses potential health risks to patients, radiation workers, and the general public if facilities are not adequately shielded. The study evaluated the performance of eight machine learning models for predicting lead-equivalent shielding thickness (mmPb). Machine learning analysis revealed Decision Tree achieved the highest test R² (0.731) with Unshielded Kerma contributing 97-99% of predictive importance, while Logistic Regression achieved perfect classification (Accuracy ꞊ 1.0) in distinguishing safe from at-risk barriers. With a test R² of 0.731 and the lowest RMSE of 0.159 mmPb, the Decision Tree model is the best performer. There are non-linear relationships in the data since all tree-based models perform better than linear models. Gradient Boosting shows the best cross-validation performance (CV R² = 0.948 ± 0.040), indicating better generalization. The performance of linear models is moderate (R² = 0.61-0.65), however KNN and SVR perform badly, suggesting that they are not appropriate for this application. With an accuracy that is clinically acceptable and forecasts that are usually within 0.16 mmPb of actual values, the Decision Tree's R² of 0.731 indicates that it accounts for 73.1% of the variance in necessary shielding. Machine learning models effectively predict shielding requirements and identify at-risk barriers.