Automatic chest computed tomography image noise quantification using deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Image noise is a critical factor affecting diagnostic accuracy in clinical computed tomography (CT).
• Traditional noise quantification methods often require multiple scans or specific homogeneous regions, limiting their clinical applicability.
• Objective and efficient noise assessment is crucial for maintaining high image quality.

Purpose of the Study:

• To develop a deep learning (DL) method for accurate noise quantification in clinical chest CT images.
• To enable noise estimation from a single CT scan, eliminating the need for repeated scans or homogeneous tissue assumptions.
• To create a tool for objective CT image quality evaluation and protocol optimization.

Main Methods:

• A convolutional neural network (CNN) was trained on a large phantom CT dataset (9240 slices) with varying dose levels and reconstruction methods.
• The CNN was designed to output local image noise standard deviations (SD) from a single CT scan input.
• The trained model was validated on diverse phantom data and subsequently applied to publicly available clinical chest CT images.

Main Results:

• The DL-based noise quantification showed strong agreement with ground truth values in phantom studies (errors < 5 HU).
• Noise SD maps generated by the CNN visually and numerically correlated well with reference estimates in clinical images.
• The method successfully produced noise SD maps for clinical data, even in areas with complex tissue interfaces and textures.

Conclusions:

• Deep learning enables feasible prediction of local noise magnitudes in CT images without repeated scanning.
• The developed DL model, trained on phantom data, effectively generalizes to clinical chest CT images.
• Automatic DL-based noise mapping offers a promising tool for objective CT image quality assessment and optimizing imaging protocols.