RHLNet: Robust Hybrid Loss-based Network for Low-Dose CT Image Denoising

• Published in: IEEE transactions on instrumentation and measurement 2024, p.1-1
• Main Authors: Saidulu, Naragoni, Muduli, Priya Ranjan, Dasgupta, Anirban
• Format: Article
• Language: English
• Subjects: Computed tomography; Convolution; Deep and shallow networks; Ensemble Learning; Huber reconstruction loss; Image reconstruction; Lattice residual block; Lattices; Linear programming; Low-dose CT denoising; Noise; Noise reduction
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2024.3403187

Low-dose computed tomography (LDCT) is a viable solution for clinical diagnosis despite noise and artifacts degrading diagnostic quality. Additionally, patients are protected from excessive X-ray radiation dose during the repeated CT scans obtained by normal-dose CT (NDCT). The data-driven deep learning algorithms with convolutional neural networks (CNN) are more effective at modeling structured CT data than random noise present. However, the aggressive denoising capability decimates structural characteristics. Also, the blurring effect may be introduced into reconstructed CT images via per-pixel-based objective functions such as mean squared error (MSE). Moreover, due to the perceptual loss computed on the feature space of the pre-trained VGG19 network, CT-specific details may be absent in perceptual features. In this study, we mitigate these problems with three novel accretions. First, we propose ensemble learning-based end-to-end low-dose CT denoising architecture using shallow and deep networks. We introduce novel lattice residual blocks (LRB) into a deep network architecture to retain long-range pixel dependency with weighted channel information. While preserving the local structure details, the three-layer shallow network accelerates the convergence of the deep network. Furthermore, we introduce Huber reconstruction loss, which combines the advantages of MSE and MAE for smooth and edge regions, respectively. We fine-tune VGG19 parameters to CT-specific data via transfer learning to avoid further loss of CT features. Finally, we include the Charbonnier loss as a structure loss term for improving visual quality. We perform extensive numerical experiments and analysis on the publicly available ''2016 NIH-AAPM-Mayo Clinic LDCT Grand Challenge" dataset.