Evaluating a Convolutional Neural Network Noise Reduction Method When Applied to CT Images Reconstructed Differently...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) are increasingly used for image denoising in medical imaging.
• The performance of these algorithms is sensitive to variations in image reconstruction parameters.

Purpose of the Study:

• To evaluate the impact of differing image reconstruction parameters on a narrowly trained CNN denoising algorithm.
• Assess the robustness of CNN denoising when applied to data not matching training conditions.

Main Methods:

• A residual CNN was trained on noise-inserted medical images with specific reconstruction settings (275 mm FOV, D30 kernel, 3 mm thickness).
• The trained CNN was tested on images reconstructed with varied parameters (100-450 mm FOV, smooth to sharp kernels, 1-5 mm thickness).

Main Results:

• Denoising efficiency decreased with smaller field of view (FOV) or smoother kernels.
• Image resolution degraded when applied to increased FOV, sharper kernels, or decreased thickness.
• Increased image thickness did not negatively affect denoising performance.

Conclusions:

• The CNN denoising algorithm's performance is significantly degraded by variations in FOV, kernel type, and image thickness.
• The algorithm is not robust to reconstruction parameter shifts from its training data.