CT Dose Optimization and Personalized Radiation Risk Assessment for Pediatric Patients using AI

Abstract

Computed Tomography (CT) is one of the most commonly used radiological exami- nations; however, dose optimization in pediatric patients poses a significant challenge due

to their higher radiosensitivity and the lack of individualized approaches in conventional methods. Therefore, this study aims to develop a machine learning-based framework to accomplish three core tasks: Regression to predict CTDIvol and DLP dose indices, Classification to stratify patients into three radiation risk categories (low, moderate, high), and Recommendation to suggest appropriate imaging protocols based on each patient’s characteristics, all in compliance with the ALARA principle (As Low As Reasonably Achievable), which seeks to minimize radiation exposure to the lowest possible level without compromising diagnostic image quality. The methodology was applied to a dataset of 359 pediatric patients, incorporating clinical variables (age, weight, body diameter) and technical parameters (mAs, kVp). Four machine learning models were compared: Random Forest, Gradient Boosting, Support Vector Machine (SVM), and Multilayer Perceptron (MLP). The results demonstrated that the Gradient Boosting model outperformed others in regression tasks, achieving an R2 of 0.9825 for CTDIvol prediction and 0.9424 for DLP prediction. Meanwhile, the MLP model achieved the best performance in radiation risk classification, with an accuracy of 87.5% and an F1-score of 0.8723. Feature importance analysis further confirmed that technical factors (particularly mAs) and body diameter are the most influential determinants of radiation dose, providing a robust foundation for clinical decision support systems based on personalized recommendations. Nevertheless, the study acknowledges certain limitations, including the relatively small sample size and the restriction of data to GE and Siemens scanners, which underscores the need for future expansion of the database and the integration of objective image quality metrics to enhance the accuracy of classification and recommendation tasks.