Fused Attention Modules Embedded in Artificial Neural Networks for Low Dose CT Denoising With Integrated Loss Functions

Abstract

<p>X-ray Computed Tomography (CT) is a non-invasive medical diagnostic tool that has raised public concerns due to the associated health risks of radiation dose to patients. Reducing the radiation dose leads to noise artifacts, making the low-dose CT images unreliable for diagnosis. Hence, low-dose computed tomography (LDCT) image reconstruction techniques have offered a new challenge in the research area. This thesis focuses on reconstructing LDCT images using deep learning techniques to provide an efficient, effective, and accurate training regimes for LDCT image denoising. A fusion of spatial and channel attention modules integrated into a dilated residual network is proposed to improve the structural details of denoised LDCT images. Further, a combination of perceptual loss, per-pixel loss, and structural dissimilarity loss is used for the optimization of the overall network. These objective functions aim to preserve structural details, avoid edge over-smoothing and enhance the image texture, respectively. Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Metrics (SSIM) are used for measuring the quantitative results. A comparative experiment was done between the proposed model and the recent denoising models, such as Block Matching and 3D Filtering (BM3D), patch Markovnian Generative Adversarial Network (patch-GAN) and dilated residual learning with edge detection (DRL-E-MP). Not only with quantitative results, but these models were also compared visually. To further strengthen the validity of the outcomes, five different CT image datasets were used. The proposed model obtained the highest PSNR/SSIM value of 34.36/0.6971 while BM3D resulted in the lowest value with 30.24/0.4461 using the chest dataset from the Mayo Clinic. Overall, the proposed network demonstrated that it could outperform state-of-the-art models.</p>