Deep Learning-Based Image Noise Quantification Framework for Computed Tomography.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Healthcare

Background:

• Accurate noise quantification is crucial for computed tomography (CT) image quality and protocol optimization.
• Traditional noise assessment methods can be time-consuming and complex.

Purpose of the Study:

• To propose and evaluate a deep learning framework, Single-scan Image Local Variance EstimatoR (SILVER), for estimating local noise levels in CT images.
• To generate pixel-wise noise maps directly from single CT scans.

Main Methods:

• Developed a U-Net convolutional neural network architecture with mean-square-error loss for noise estimation.
• Trained the model using 120,000 phantom CT images, with pixel-wise noise maps calculated from 100 replicate scans.
• Evaluated SILVER performance on both phantom and patient images, comparing results with manual measurements.

Main Results:

• SILVER accurately predicted noise maps on phantom data, closely matching calculated noise maps (RMSE <8 HU).
• On patient images, SILVER demonstrated a low average percent error of 5% compared to manual region-of-interest measurements.

Conclusions:

• The SILVER framework enables precise, pixel-wise noise level estimation directly from patient CT images.
• This accessible method, trained on phantom data, has broad applicability in CT image analysis and optimization.