Quantitative evaluation of deep convolutional neural network-based image denoising for low-dose computed tomography

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Minimizing radiation exposure in computed tomography (CT) is crucial for patient safety.
• Image noise in low-dose CT compromises diagnostic quality.
• Existing deep learning denoising methods may cause blurring or gradient loss.

Purpose of the Study:

• To compare a CNN-based denoising method (DnCNN) against other noise-reduction techniques for low-dose CT.
• To evaluate the dose-dependent performance of CNN denoising.
• To assess image quality using structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR).

Main Methods:

• Simulated low-dose CT images by adding Poisson noise and applying CT-specific modulation transfer function.
• Utilized 100 abdominal CT images from a public database, creating dose-reduction intervals down to 1/100th dose.
• Employed DnCNN for denoising and evaluated image quality metrics (SSIM, PSNR).

Main Results:

• DnCNN significantly outperformed other methods in denoising, particularly at 10% and 5% dose levels.
• The CNN model effectively reduced noise and maintained image sharpness, improving SSIM by ~10%.
• Excessive smoothing occurred under small dose-reduction conditions, indicating a need for model tailoring.

Conclusions:

• CNN-based denoising offers superior performance for low-dose CT image quality enhancement.
• Tailoring the CNN model is essential to prevent over-smoothing and optimize noise reduction.
• This approach holds promise for improving diagnostic accuracy in low-dose CT applications.