An artificial intelligence model for predicting an appropriate mAs with target exposure indicator for chest digital...

Area of Science:

• Radiological physics
• Medical imaging
• Artificial intelligence in healthcare

Background:

• Digital radiography image quality depends on multiple factors, including patient anatomy and exposure settings.
• Accurate estimation of exposure factors like current-time product (mAs) is challenging, leading to potential overexposure or underexposure.
• There is a need for predictive models to optimize mAs for diagnostic image quality.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) model for predicting optimal mAs in chest digital radiography.
• To utilize a target exposure indicator, derived from phantom studies, for defining appropriate exposure levels.
• To validate the ML model's performance in a human study cohort.

Main Methods:

• An anthropomorphic chest phantom was used to establish a target exposure indicator (average Reached Exposure [REX] = 355.6).
• A dataset of 1,000 patients undergoing chest radiography was collected, recording parameters like chest thickness, weight, and exposure factors (mAs, REX).
• Five ML models were trained and validated using 80% training and 20% testing data splits, with the Artificial Neural Network (ANN) identified as the most suitable.

Main Results:

• The ANN model demonstrated high accuracy in predicting both REX and mAs values.
• Predicted mAs values were, on average, 10% lower in overexposed cases (REX > 355.6) and 8% higher in underexposed cases (REX < 355.6) compared to Automatic Exposure Control (AEC).
• The model showed potential for further mAs reduction by adjusting the target REX, indicating dose optimization possibilities.

Conclusions:

• Machine learning provides a feasible approach for developing AI models to predict appropriate mAs in chest digital radiography.
• The developed ML model can help optimize image quality and reduce patient radiation dose by estimating precise exposure factors.
• This AI-driven prediction of mAs represents a significant advancement in enhancing diagnostic accuracy and safety in digital radiography.