AI-Driven Prediction of Chest CT Radiation Doses: Establishing BMI-Based Diagnostic Reference Levels and
Zuhal Y Hamd1, Mohamed Abuzaid2, Mohamed Alharbi3
1Department of Radiological Sciences, College of Health and Rehabilitation Sciences, Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh 11671, Saudi Arabia.
Tomography (Ann Arbor, Mich.)
|May 26, 2026
Summary
New body mass index (BMI)-stratified diagnostic reference levels (DRLs) for adult chest CT scans improve radiation dose management. AI models predict radiation dose, enabling prospective dose governance and protocol optimization.
Area of Science:
- Radiological Physics
- Medical Imaging
- Radiation Dosimetry
Background:
- Computed tomography (CT) scans of the chest significantly contribute to population radiation exposure.
- Traditional diagnostic reference levels (DRLs) lack personalization for body size and are applied retrospectively.
- Individual variability in patient body habitus necessitates a more tailored approach to radiation dose assessment in chest CT.
Purpose of the Study:
- To evaluate radiation dose patterns in adult chest CT examinations.
- To establish body mass index (BMI)-stratified local DRLs for chest CT.
- To develop artificial intelligence (AI)-assisted models for predicting radiation dose before scanning.
Main Methods:
- Analysis of consecutive adult chest CT scans, extracting dose indices (CTDIvol, DLP) and patient factors (BMI, weight, height, age, sex).
- Definition of DRLs using the 75th percentile, stratified by BMI categories (underweight, normal, overweight, obese).
- Application of supervised learning algorithms (Random Forest, Gradient Boosting) to predict CTDIvol and DLP using routine variables.
Main Results:
- BMI-stratified DRLs showed a monotonic increase with BMI categories, with significant differences observed for both DLP and CTDIvol across groups (p < 0.001).
- Radiation dose indices (DLP) moderately correlated with weight and BMI (r ≈ 0.54-0.56), while height was not a significant predictor.
- Developed predictive models demonstrated high performance (R² up to ~0.79 for CTDIvol, ~0.77 for DLP), allowing for pre-acquisition dose comparison against BMI-matched DRLs.
Conclusions:
- BMI-stratified DRLs offer clinically relevant benchmarks, overcoming limitations of pooled DRLs.
- AI-driven prediction of radiation dose enables prospective dose governance at the point of care.
- This framework facilitates optimized radiation dose management and protocol refinement in chest CT imaging.
