[Paper Review] Deep Learning for Low-Dose CT Denoising

Maryam Gholizadeh-Ansari, Javad Alirezaie+1 more|arXiv (Cornell University)|Feb 25, 2019

Medical Imaging Techniques and Applications38 references4 citations

TL;DR

This paper proposes a deep learning framework for low-dose CT denoising that combines dilated convolutions, residual learning, a non-trainable edge detection layer, and a hybrid loss function of mean squared error (MSE) and perceptual loss. The method significantly improves image quality by preserving fine textures and edges while reducing over-smoothing and grid-like artifacts, achieving state-of-the-art performance on simulated and real low-dose CT datasets.

ABSTRACT

Low-dose CT denoising is a challenging task that has been studied by many researchers. Some studies have used deep neural networks to improve the quality of low-dose CT images and achieved fruitful results. In this paper, we propose a deep neural network that uses dilated convolutions with different dilation rates instead of standard convolution helping to capture more contextual information in fewer layers. Also, we have employed residual learning by creating shortcut connections to transmit image information from the early layers to later ones. To further improve the performance of the network, we have introduced a non-trainable edge detection layer that extracts edges in horizontal, vertical, and diagonal directions. Finally, we demonstrate that optimizing the network by a combination of mean-square error loss and perceptual loss preserves many structural details in the CT image. This objective function does not suffer from over smoothing and blurring effects caused by per-pixel loss and grid-like artifacts resulting from perceptual loss. The experiments show that each modification to the network improves the outcome while only minimally changing the complexity of the network.

Motivation & Objective

• To address the challenge of noise in low-dose CT scans, which increases radiation risk and reduces diagnostic accuracy.
• To develop a fast, vendor-independent image space denoising method that does not require access to sinogram data.
• To improve image quality by preserving structural details and edges while minimizing over-smoothing and artifacts.
• To combine perceptual loss with MSE loss to balance perceptual fidelity and pixel-level accuracy.
• To demonstrate the effectiveness of architectural innovations—dilated convolutions, residual connections, and edge detection—on real and simulated low-dose CT data.

Proposed method

• Uses dilated convolutions with multiple dilation rates to expand感受 field and capture more contextual information without increasing network depth.
• Employs residual learning with shortcut connections to facilitate training and preserve feature representation across layers.
• Introduces a non-trainable edge detection layer inspired by the Sobel operator to detect horizontal, vertical, and diagonal edges, enhancing boundary preservation.
• Optimizes the network using a hybrid loss function combining mean squared error (MSE) for pixel-level accuracy and perceptual loss for structural similarity to the ground truth.
• Trains the network end-to-end on paired low-dose and normal-dose CT images, with the objective of reconstructing high-quality images from low-dose inputs.
• Evaluates the model on three datasets: simulated lung, real piglet, and phantom thoracic scans, using PSNR and SSIM for quantitative comparison.

Experimental results

Research questions

• RQ1 Can dilated convolutions and residual learning improve low-dose CT denoising performance while maintaining computational efficiency?
• RQ2 Does integrating a non-trainable edge detection layer enhance edge preservation in low-dose CT images?
• RQ3 Can a combined MSE and perceptual loss function mitigate over-smoothing from MSE and grid-like artifacts from perceptual loss?
• RQ4 How does the proposed method compare to state-of-the-art techniques like BM3D, CNN200, and prior-based models on real and simulated data?
• RQ5 To what extent do architectural modifications improve visual and quantitative image quality in low-dose CT denoising?

Key findings

• The proposed DRL-E-MP model achieved the highest PSNR (44.12) and SSIM (0.9807) on the Piglet dataset, outperforming BM3D, CNN200, and prior-based methods.
• On the Thoracic phantom dataset, DRL-E-MP achieved a PSNR of 31.50 and SSIM of 0.6381, significantly outperforming models using only MSE or perceptual loss.
• Visual results showed that DRL-E-MP preserved fine textures and edges better than models trained with MSE alone, which caused over-smoothing, and perceptual loss alone, which introduced grid-like artifacts.
• The edge detection layer improved edge sharpness, as confirmed by visual comparison between DRL and DRL-E-M on the Thoracic dataset.
• The hybrid loss function (MSE + perceptual) achieved the best balance, producing visually superior images without sacrificing quantitative metrics.
• Each architectural component—dilated convolutions, residual learning, edge detection, and combined loss—contributed incrementally to performance improvement with minimal increase in model complexity.