Potential of a machine-learning model for dose optimization in CT quality assurance

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Quality assurance in CT imaging is crucial for optimizing radiation dose.
• Retrospective analysis of CT dose data can be complex and time-consuming.

Purpose of the Study:

• To evaluate machine learning (ML) for detecting chest CT examinations with dose optimization potential.
• To assess ML's utility in simplifying CT quality assurance processes.

Main Methods:

• A feed-forward neural network was trained on 3,199 chest CT examinations to predict volumetric computed tomography dose index (CTDIvol).
• Model inputs included scanner, study description, protocol, patient demographics, and water-equivalent diameter (Dw).
• A separate validation set of 100 CTs was reviewed by radiologists to define an optimal ML model cutoff.

Main Results:

• The ML model achieved a root mean-squared error (RMSE) of 1.52 on the validation dataset.
• Scanner and Dw were identified as the most influential features for dose prediction.
• The ML model flagged 8% of cases as suboptimal, correctly identifying all 7% flagged by radiologists, with one additional case flagged by the model alone.

Conclusions:

• Machine learning can effectively detect chest CT examinations with dose optimization potential.
• ML integration can enhance retrospective CT dose data analysis and quality assurance.
• While ML is a valuable tool, final human review by radiologists or physicists is necessary to confirm findings.