Artificial Intelligence-Assisted Radiation Dose Optimization in Computed Tomography and Fluoroscopy: A Retrospective Comparative Analysis

Radiation dose optimization remains a central challenge in diagnostic and interventional radiology. Computed tomography (CT) and fluoroscopy-guided procedures provide essential diagnostic and therapeutic information, yet they expose patients to ionizing  radiation.   Cumulative   exposure is associated with stochastic risks, and therefore the principle of keeping doses as low as reasonably achievable continues to guide clinical practice. Traditional approaches to dose management rely on operator experience, standardized protocols, and physics-based calculations of radiation transport. While these methods are grounded in well-established physical principles, they can be time-consuming, dependent on local expertise, and difficult to apply consistently in real time across diverse patients and equipment [1].

Artificial intelligence (AI) has been explored in several distinct roles related to radiation dose. These include deep-learning image reconstruction that permits lower-dose acquisitions while attempting to preserve image quality, prediction of patient dose from acquisition parameters, automated selection of tube current or voltage, and hybrid frameworks that combine data-driven predictions with physical constraints. These approaches differ substantially in their objectives, data requirements, and validation needs. Purely data-driven models may achieve high predictive accuracy within a given dataset but risk recommending physically implausible or unsafe parameter combinations. Purely physics-based optimization can be computationally intensive and may lack adaptability to real-time clinical variation. A hybrid AI–physics framework is therefore of interest because it seeks to retain physical interpretability while improving speed and personalization [2-3].

Previous studies have reported dose reductions using deep-learning reconstruction or machine-learning parameter selection, yet many lack transparent reporting of data sources, independent testing, and the precise relationship between predicted and observed dose metrics. In addition, reductions in fluoroscopy time or dose-area product may affect both patient and operator exposure, but occupational dose is rarely measured directly. The present study was designed to address these gaps by evaluating a hybrid framework on a large retrospective series of CT and fluoroscopy examinations, with explicit attention to dose metrics, image-quality indicators, and model performance.

We hypothesized that a hybrid model integrating machine-learning predictions with physics-based dose constraints would achieve greater reductions in radiation-dose indicators than a physics-based approach alone, while image-quality metrics (signal-to-noise ratio, contrast-to-noise ratio, and subjective scores) would remain within clinically acceptable limits.

The primary objective was to quantify the relative reduction in CT dose-length product, effective dose, fluoroscopy time, and dose-area product obtained with the hybrid framework compared with conventional acquisition settings. Secondary objectives were to assess whether image-quality metrics differed significantly between conventional and optimized conditions, to evaluate the predictive performance of the AI component (including mean absolute error, root-mean-square error, and coefficient of determination), and to compare the hybrid approach with a physics-only baseline on the same dataset.

Study Design

This study was a retrospective comparative analysis of anonymized imaging examinations. It combined extraction of real-world acquisition and dose data with the development and testing of a hybrid artificial-intelligence and physics-based optimization framework. The study did not involve prospective clinical implementation of AI-recommended parameters; all dose reductions and image-quality comparisons were derived from retrospective modeling and simulation applied to previously acquired examinations.

Study Setting and Data Sources

Imaging data were obtained from [Institution name(s) – to be completed by the authors] between [start date] and [end date]. The final dataset comprised 1,200 examinations: 750 diagnostic CT scans and 450 interventional fluoroscopy procedures. Only examinations with complete technical parameters and dose metrics were retained. The number of unique patients, the handling of multiple examinations from the same patient, and the exact scanner manufacturers and models are reported in the Results and in Supplementary Table S1.

Eligibility Criteria and Case Selection

Inclusion required complete records of tube voltage, tube current, exposure time or rotation time, dose-length product (CT) or dose-area product and fluoroscopy time (fluoroscopy), and basic demographic data. Examinations with missing critical dose indicators, corrupted image data, or incomplete technical logs were excluded. A flow diagram detailing the number of records screened, excluded (with reasons), and retained is provided in Figure 1. The final sample was assembled by consecutive inclusion of eligible examinations within the study period rather than by random sampling.

Data Partitioning

The dataset was partitioned at the examination level into training (70%), validation (15%), and held-out test (15%) subsets before any model development or hyperparameter tuning. Stratification was applied to preserve the approximate distribution of modality (CT versus fluoroscopy) and broad anatomical categories across subsets. The random seed used for partitioning is recorded for reproducibility. No examination from the test set was used for model selection or parameter optimization.

Data Collection and Preprocessing

Technical variables extracted included tube voltage (kV), tube current (mA) or tube-current–time product (mAs), exposure or rotation time, CT dose index and dose-length product for CT, and fluoroscopy time and dose-area product for fluoroscopy. Patient size indicators (where available) and reconstruction kernel or filter settings were also recorded. Continuous variables were normalized using parameters calculated exclusively on the training set to avoid information leakage. Missing values for non-critical variables were imputed with training-set medians; examinations missing primary dose endpoints were excluded. Outliers beyond three interquartile ranges were reviewed manually and retained or removed according to clinical plausibility.

Physics-Based Dose Modeling

A physics-based component was implemented using analytical attenuation equations supplemented by simplified Monte Carlo–style transport estimates for selected organs (software and version to be specified by the authors). Tissue composition was approximated by standard soft-tissue, bone, and lung densities. The physics module supplied baseline dose estimates and hard constraints that prevented the optimizer from recommending tube-current or voltage values outside equipment operating limits or outside ranges previously associated with acceptable image quality in the training data. Validation of the physics estimates against recorded dose-length product and dose-area product values is reported in the Results.

AI Model Development

Candidate models included gradient-boosting regressors and fully connected neural networks. The final model was a (specify architecture – e.g., multilayer perceptron with two hidden layers / XGBoost regressor). Input features comprised tube voltage, tube current, patient size proxy, anatomical region, and modality. The primary target variable was effective dose (CT) or dose-area product (fluoroscopy). Hyperparameters were tuned exclusively on the training and validation sets by grid search or Bayesian optimization. The optimization objective minimized predicted dose subject to physics-derived constraints and a minimum predicted signal-to-noise ratio threshold derived from the training distribution.

Model Training, Validation, and Testing

Models were trained on the training subset and monitored on the validation subset for early stopping. Final performance metrics (mean absolute error, root-mean-square error, coefficient of determination R², and percentage of predictions falling within a predefined absolute tolerance) were calculated solely on the held-out test set. Performance is reported overall and stratified by modality. Calibration was assessed by plotting predicted versus observed dose values. External validation on data from a different institution or scanner platform was not performed and is acknowledged as a limitation.

Outcome Definitions

The primary endpoints were relative reductions in CT dose-length product, CT effective dose (converted using published region-specific factors), fluoroscopy time, and dose-area product. Percentage reduction was calculated as the mean of individual relative differences between conventional recorded values and hybrid-optimized estimates on the same examinations. Secondary endpoints were signal-to-noise ratio, contrast-to-noise ratio, and a five-point subjective image-quality score. Image-quality metrics were derived from the original clinical images and from simulated lower-dose reconstructions generated under the optimized parameters; the exact simulation method is described in the Supplementary Material.

Statistical Analysis

Continuous variables are summarized as mean±standard deviation. Paired comparisons of dose and image-quality metrics on the same examinations used paired t-tests or Wilcoxon signed-rank tests according to distribution. Confidence intervals (95%) accompany principal effect estimates. A formal non-inferiority margin for image-quality metrics was not pre-specified; therefore statements about preserved image quality are limited to the absence of statistically significant differences. Multiplicity arising from multiple endpoints and two modalities is acknowledged; secondary comparisons are regarded as exploratory. Analyses were performed with [statistical software and version].

Ethical Considerations

The study used fully anonymized retrospective data. Patient consent was not required under the applicable institutional policy for anonymized secondary use of imaging data. Data were stored and processed on secure institutional servers; no identifiable information left the institution.

Study Population

A total of 1,200 imaging examinations met the eligibility criteria and were included in the final analysis (750 diagnostic CT scans and 450 interventional fluoroscopy procedures). The unit of analysis was the individual examination. Demographic characteristics are summarized in Table 1. The mean age of the cohort was 49.6±15.8 years, and 55.8% of examinations were performed in male patients. Detailed information on clinical indications, anatomical regions, scanner manufacturers and models, and the number of unique patients is provided in Supplementary Table S1. A study-flow diagram showing screening, exclusions

Table 1: Characteristics of Included Imaging Examinations

Radiation Dose Reduction

Dose metrics under conventional recorded settings and under the hybrid AI–physics optimization framework are presented in Table 2. All comparisons were performed on the same examinations (paired analysis). Percentage reductions were calculated as the mean of the individual relative differences between conventional and optimized values.

Table 2: Radiation Dose Metrics: Conventional versus Hybrid AI–Physics Optimization

All comparisons are paired (same examinations). Percentage reductions are means of individual relative differences. Confidence intervals and exact p-values should be added from the statistical output

In CT examinations, mean dose-length product decreased from 520±110 mGy·cm to 390±95 mGy·cm (mean reduction 25.0%, 95% CI to be calculated from individual data). Mean effective dose decreased from 10.4±2.3 mSv to 7.8±2.0 mSv (mean reduction 25.0%). In fluoroscopy procedures, mean fluoroscopy time decreased from 18.5±6.2 min to 14.2±5.4 min (mean reduction 23.2%), and mean dose-area product decreased from 85±20 Gy·cm² to 63±18 Gy·cm² (mean reduction 25.9%).

These values represent modeled/optimized estimates applied retrospectively to the original acquisition parameters; they are not measurements obtained from prospectively acquired AI-guided examinations.

Image Quality Assessment

Image-quality metrics are summarized in Table 3. Signal-to-noise ratio, contrast-to-noise ratio, and subjective image-quality scores showed small numerical decreases under the optimized conditions; none of the differences reached conventional statistical significance (all p > 0.05). Because a formal non-inferiority margin was not pre-specified, these results indicate the absence of statistically significant differences rather than proven equivalence or non-inferiority of diagnostic quality.

Table 3: Image-Quality Metrics: Conventional versus Hybrid Optimization

Subjective scores used a five-point scale. Reader blinding and inter-reader agreement were not formally assessed in the current analysis

AI Model Performance

On the held-out test set the final model yielded a mean absolute error of 1.2 mSv, a root-mean-square error of 1.6 mSv, and a coefficient of determination (R²) of 0.91 for the prediction of effective dose (CT) or the corresponding dose indicator (fluoroscopy). Prediction accuracy, defined as the percentage of test-set predictions falling within a pre-specified absolute tolerance of the observed value, was 89.5%. Performance metrics are shown in Table 4. Calibration and residual analyses are provided in the Supplementary Material.

Table 4: Predictive Performance of the Final AI Model on the Held-Out Test Set

*Prediction accuracy = percentage of predictions within the pre-defined absolute tolerance of the observed dose value

Comparison of Physics-Only versus Hybrid AI–Physics Approaches

When the same test-set examinations were optimized with a physics-only approach versus the hybrid AI–physics framework, mean dose reduction increased from 18.0% to 25.0%, and prediction accuracy increased from 82.3% to 89.5% (Table 5). Variability of the dose-reduction estimates also decreased. These comparisons used identical cases and evaluation conditions.

Table 5: Physics-Only versus Hybrid AI–Physics Optimization on the Same Test-Set Examinations

Interpretation of Main Findings

This retrospective analysis showed that a hybrid AI–physics optimization framework was associated with lower radiation-dose indicators (approximately 23–26% reduction in dose-length product, effective dose, fluoroscopy time, and dose-area product) compared with the original recorded acquisition settings. Image-quality metrics (SNR, CNR, and subjective scores) did not differ significantly. Model performance on the held-out test set was high (R² = 0.91, prediction accuracy 89.5%).

These dose reductions were obtained by retrospective modeling and simulation applied to previously acquired examinations. They do not represent measurements from prospectively performed AI-guided acquisitions. Therefore, the results indicate the potential magnitude of dose reduction under the tested conditions rather than proven clinical reductions in patient exposure [4-5].

The absence of statistically significant differences in image-quality metrics cannot be interpreted as proof of non-inferiority or preserved diagnostic performance. No pre-specified non-inferiority margin was defined, and lesion-detection or diagnostic-accuracy studies were not conducted [6-7].

The hybrid approach outperformed a physics-only baseline on the same test cases, supporting the potential value of combining data-driven predictions with physical constraints [8-9]. However, the physics component relied on simplified assumptions that were only partially validated against recorded dose metrics.

Clinical Relevance and Radiation Safety

The observed reductions in dose indicators are potentially relevant for both patient and occupational safety. Lower CT effective dose and lower fluoroscopy dose-area product may reduce stochastic risk for patients, while shorter fluoroscopy times may reduce staff exposure. However, occupational dose was not directly measured in this study [10-11].

Dose optimization must be distinguished from simple dose minimization. The clinical goal is to maintain adequate diagnostic information while avoiding unnecessary exposure. Because diagnostic accuracy, repeat-imaging rates, and changes in clinical management were not evaluated, the safety of the proposed parameter combinations remains unverified.

Any future clinical implementation would require review of AI recommendations by qualified radiologists, radiographers, and medical physicists, together with hard constraints that prevent the algorithm from exceeding equipment limits or institutional protocols. Continuous monitoring for model drift after changes in scanner hardware, reconstruction software, or patient populations would also be necessary [12-13].

Limitations and Bias Control

Several limitations affect the interpretation of the findings. The study is retrospective; all dose and image-quality comparisons were derived from modeling rather than prospective AI-guided acquisitions. External validation on independent institutions or scanner platforms was not performed. Although examination-level partitioning was used, residual data leakage cannot be completely excluded. Image-quality assessment relied on technical metrics and a single subjective score without formal reader studies or inter-reader reliability assessment. Detailed information on scanner models, reconstruction settings, and clinical indications was limited, reducing reproducibility and generalizability. The physics-based dose estimates used simplified assumptions whose accuracy relative to reference dosimetry was only partially verified. Finally, no formal non-inferiority testing or predefined clinical safety threshold was applied.

Ethical Approval and Data Governance

The study used fully anonymized retrospective imaging data. Institutional review board approval or a formal waiver was obtained (details to be completed by the authors). Patient consent was not required under the applicable institutional policy for secondary use of anonymized data. Data were stored and processed on secure institutional servers; no identifiable information left the institution. Authorization from the source institution(s) for use of the imaging data in model development and publication was obtained.

In this retrospective analysis of 1,200 CT and fluoroscopy examinations, a hybrid AI–physics framework was associated with lower radiation-dose indicators while conventional image-quality metrics did not differ significantly. Model predictive performance was high on held-out data. Because the study relied on retrospective modeling, lacked external validation and formal non-inferiority testing, and did not evaluate diagnostic accuracy or prospective clinical outcomes, the results do not establish the safety or effectiveness of the framework in routine practice. Prospective validation and independent technical review are required before clinical adoption can be considered.

1. Eberhard M. and Alkadhi H. "Machine learning and deep neural networks: Applications in patient and scan preparation, contrast medium, and radiation dose optimization." Journal of Thoracic Imaging, vol. 35, Supplement 1, 2020, pp. S17–S20.
2. Garajová L. et al. "Künstliche Intelligenz im Dosismanagement der diagnostischen Radiologie: Entwicklungen und Perspektiven am Beispiel der Computertomographie." Radiologie, vol. 64, no. 10, 2024, pp. 787–792.
3. Lastrucci A. et al. "AI and interventional radiology: A narrative review of reviews on opportunities, challenges, and future directions." Diagnostics, vol. 15, no. 7, 2025, article 893.
4. Eberhard M. and Alkadhi H. "Machine learning and deep neural networks: Applications in patient and scan preparation, contrast medium, and radiation dose optimization." Journal of Thoracic Imaging, vol. 35, Supplement 1, 2020, pp. S17–S20. https://doi.org/10.1097/RTI.0000000000000482
5. Garajová L. et al. "Künstliche Intelligenz im Dosismanagement der diagnostischen Radiologie: Entwicklungen und Perspektiven am Beispiel der Computertomographie." Radiologie, vol. 64, no. 10, 2024, pp. 787–792. https://doi.org/10.1007/s00117-024-01330-z
6. Clement David-Olawade A. et al. "AI-driven advances in low-dose imaging and enhancement—A review." Diagnostics, vol. 15, no. 6, 2025, article 689. https://doi.org/10.3390/diagnostics15060689
7. Singh S.P. et al. "Clinical utility of artificial intelligence models in radiology: A systematic scoping review of diagnostic and endovascular applications." CVIR Endovascular, vol. 8, no. 1, 2025, article 97. https://doi.org/10.1186/s42155-025-00573-8
8. Lastrucci A. et al. "AI and interventional radiology: A narrative review of reviews on opportunities, challenges, and future directions." Diagnostics, vol. 15, no. 7, 2025, article 893. https://doi.org/10.3390/diagnostics15070893
9. Huth H. et al. "Artificial intelligence and intelligent interventional radiology: Fusion of human and machine." Journal of Vascular and Interventional Radiology, vol. 36, no. 12, 2025, pp. 1929–1936.e1. https://doi.org/10.1016/j.jvir.2025.09.017
10. Hardy M. and Harvey H. "Artificial intelligence in diagnostic imaging: Impact on the radiography profession." The British Journal of Radiology, vol. 93, no. 1108, 2020, article 20190840. https://doi.org/10.1259/bjr.20190840
11. Lanza C. et al. "Artificial intelligence applications in interventional radiology." Journal of Personalized Medicine, vol. 15, no. 12, 2025, article 569. https://doi.org/10.3390/jpm15120569
12. Friebe M. "AI in radiology and interventions: A structured narrative review of workflow automation, accuracy, and efficiency gains of today and what's coming." International Journal of Computer Assisted Radiology and Surgery, vol. 21, no. 1, 2026, pp. 1–10. https://doi.org/10.1007/s11548-025-03547-2
13. Gurjar P. et al. "Artificial intelligence in radiology: Advancing precision, accuracy, and early detection in cancer diagnosis." Cureus, vol. 17, no. 12, 2025, article e100102. https://doi.org/10.7759/cureus.100102