Structure-sensitive Multi-scale Deep Neural Network for Low-Dose CT Denoising (1805.00587v3)

Abstract: Computed tomography (CT) is a popular medical imaging modality in clinical applications. At the same time, the x-ray radiation dose associated with CT scans raises public concerns due to its potential risks to the patients. Over the past years, major efforts have been dedicated to the development of Low-Dose CT (LDCT) methods. However, the radiation dose reduction compromises the signal-to-noise ratio (SNR), leading to strong noise and artifacts that down-grade CT image quality. In this paper, we propose a novel 3D noise reduction method, called Structure-sensitive Multi-scale Generative Adversarial Net (SMGAN), to improve the LDCT image quality. Specifically, we incorporate three-dimensional (3D) volumetric information to improve the image quality. Also, different loss functions for training denoising models are investigated. Experiments show that the proposed method can effectively preserve structural and texture information from normal-dose CT (NDCT) images, and significantly suppress noise and artifacts. Qualitative visual assessments by three experienced radiologists demonstrate that the proposed method retrieves more detailed information, and outperforms competing methods.

• The paper presents SMGAN, a novel multi-scale GAN framework that integrates 3D CNNs and a structurally-sensitive loss to effectively denoise low-dose CT images.
• The methodology combines adversarial, structural similarity, and pixel-wise loss functions to balance noise suppression with the preservation of textural and structural details.
• Experimental results demonstrate significant gains in PSNR and SSIM, outperforming existing methods and enhancing diagnostic reliability.

Structurally-sensitive Multi-scale Deep Neural Network for Low-Dose CT Denoising

The paper "Structurally-sensitive Multi-scale Deep Neural Network for Low-Dose CT Denoising" addresses the challenge of denoising low-dose computed tomography (LDCT) images by introducing a novel deep learning-based solution. The concern over ionizing radiation associated with CT scans has prompted efforts to reduce the radiation dose, yet such reductions come at the cost of degrading image quality due to increased noise. This paper proposes the Structurally-sensitive Multi-scale Generative Adversarial Net (SMGAN), a method leveraging 3D convolutional neural networks and adversarial learning to improve LDCT imaging.

The proposed model employs a multi-scale approach to incorporate volumetric information, aiming to maintain the structural integrity of images while suppressing artifacts. Different loss functions are systematically explored, leading to a novel structural loss function that combines adversarial, structural similarity, and pixel-wise loss, thereby optimizing image realism and noise reduction. Results from experiments demonstrate superior performance in preserving textural details and reducing artifacts compared to existing LDCT denoising techniques. These advancements are substantiated via qualitative assessments by experienced radiologists, who reveal that the SMGAN method retrieves significant clinical information and surpasses benchmark algorithms.

The structure of SMGAN includes three essential components: a generator network that employs a 3D CNN to generate improved images, a sensitive loss function designed to enhance denoising quality, and a discriminator network, implemented with 2.5D filters for computational efficiency. This configuration facilitates advanced noise reduction while preserving critical structural features, promoting diagnostic accuracy.

Several key advancements in optimization are discussed, notably the use of Wasserstein distance for stable adversarial training and a structurally-sensitive loss function that combines L_1 and MS-SSIM losses. These innovations enable the network to capture subtle structural nuances in a volumetric context. Comparisons with state-of-the-art denoising methods, including CNN-L2, RED-CNN, BM3D, and WGAN-VGG, highlight the efficacy of SMGAN in maintaining a balance between noise suppression and structural preservation.

The strong numerical results—such as the improvements in PSNR and SSIM scores—demonstrate the capability of SMGAN to elevate LDCT image quality to be diagnostically acceptable while minimizing radiation exposure. This paper's significant contributions entail the integration of 3D spatial information into GAN frameworks, providing a promising direction for medical imaging enhancements.

The implications of this research are manifold: practically, it offers a feasible method for imaging centers to reduce radiation doses without conceding diagnostic reliability; theoretically, it underscores the potential of adversarial networks combined with structure-aware loss functions. Future developments may refine the architectural components of SMGAN, optimize hyperparameters for various CT modalities or extend its applications beyond CT to other imaging domains like MRI or PET.

In summary, this paper presents a comprehensive approach to addressing the trade-off between radiation dose and image quality in CT imaging, proposing an innovative deep learning model that could significantly impact clinical practices and guide further advancement in AI-driven image enhancement technologies.