Deep Learning for Low-Dose CT Denoising Using Perceptual Loss and Edge Detection Layer.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Low-dose computed tomography (CT) imaging is crucial for reducing radiation exposure.
• Image noise in low-dose CT significantly degrades image quality, posing challenges for accurate diagnosis.
• Deep learning methods have shown promise in CT image denoising.

Purpose of the Study:

• To develop an advanced deep neural network for effective low-dose CT denoising.
• To enhance the preservation of structural details and reduce artifacts in CT images.
• To minimize the increase in computational complexity while improving denoising performance.

Main Methods:

• Utilizing dilated convolutions with varying dilation rates to capture extensive contextual information efficiently.
• Implementing residual learning through shortcut connections to facilitate information flow across network layers.
• Incorporating a non-trainable edge detection layer for extracting multi-directional edges (horizontal, vertical, diagonal).
• Optimizing the network using a combined loss function of mean-square error and perceptual loss.

Main Results:

• The proposed network effectively reduces noise in low-dose CT images.
• The combination of dilated convolutions and residual learning enhances contextual information capture and feature propagation.
• The edge detection layer aids in preserving critical structural details.
• The hybrid loss function mitigates over-smoothing, blurring, and grid-like artifacts.
• Each network modification demonstrated performance improvements with minimal impact on complexity.

Conclusions:

• The developed deep neural network offers a significant advancement in low-dose CT denoising.
• The proposed architectural modifications and optimization strategy lead to superior image quality preservation.
• This approach holds potential for improving diagnostic accuracy in low-dose CT examinations.