Chapter 5 - Medical image denoising using convolutional neural networks

Abstract

Medical image analysis involves correct interpretation of image data for the proper diagnosis to decide further course of treatment. There are different medical imaging modalities, such as magnetic resonance imaging, computed tomography, ultrasound, positron emission tomography, etc. Each imaging modality is used for the diagnosis of a particular disease according to its features. The image denoising is an important step of the medical image pre-processing part for accurate interpretation of image data. There are different image denoising techniques, such as transform domain-based, fuzzy logic-based, spatial domain-based, optimization-based, and machine learning-based methods. Out of these techniques, machine learning-based convolutional neural networks (CNNs) are widely being used for image denoising applications. The auto extraction of image features by the convolutional operation is the core functionality of CNNs. Moreover, CNN's design is appropriate for image data, and the concept of weight sharing reduces computational complexity as well. This chapter gives an overview of various medical image modalities and corresponding denoising CNN architectures. It has been observed that denoising CNN (DnCNN) Gaussian denoiser is the benchmark model for the design of medical imaging denoisers. The basic design unit is convolution, batch normalization, and rectified linear units. The 3D-parallel RicianNet for Rician Noise removal in MRI, two-stage residual CNN, and complex CNN medical image denoiser for CT images, mixed attention-based residual U-Net for speckle removal in ultrasound, and unsupervised learning with image priors for PET images have outperformed other state-of-the-art methods.