Multi-Scale Dilation With Residual Fused Attention Network for Low Dose CT Noise Artifact Reductions

Medical Imaging has been a growing topic in computer vision for its life-saving applications. Medical professionals rely on good quality images of their medical scans in order to correctly identify tumors or other anomalies. Recent studies have shown that Computed Tomography (CT) scans have great results using high radiations for their X-Rays. CT Scans produce more than half the radiation exposure from medical use, which results in problems for long-term use of these expensive machines. Some solutions have involved reducing the radiation dose; however, that leads to noise artifacts making the Low-Dose Computed Tomography (LDCT) images unreliable for diagnosis. In this study, a Multi-scale Dilation with Residual Fused Attention (MD-RFA) deep neural network is proposed, more specifically a network with an integration with a multi-scale feature mapping, spatial- and channel-attention module to enhance the quality of LDCT images. Further, the multiscale image mapping uses a series of dilated convolution layers, which promotes the model to capture hierarchy features of different scales. The attention modules are combined in a parallel connection and are described as boosting attention fusion blocks (BAFB) that are then stacked on top of one another, creating a residual connection known as boosting module groups (BMG). The network is designed through a series of experiments that evaluate how certain modifications affect performance. The model is optimized using an integration of mean-squared error (MSE) loss, perceptual loss via ResNet50V2-net, and Structural Dissimilarity (DSSIM) loss. Through an ablation experiment, these functions show that it could effectively prevent edge over-smoothing, improve image texture, and preserve structural details. Finally, comparative experiments show that the proposed network outperforms state-of-the-art denoising models such as Block Matching and 3D Filtering (BM3D), Markovian-based patch generative adversarial network (patch-GAN) Dilated Residual network with Edge detection (DRL-E-MP), and Fused Attention Module with Dilated Residual Learning (FAM-DRL).