Recently, deep learning-based convolutional neural networks method for image super-resolution has achieved remarkable performance in various fields including security surveillance, satellite imaging, and medical image enhancement. Although these approaches obtained improved performance in medical images, existing works only used a pre-processing step and hand-designed filter methods to improve the quality of medical images. Pre-processing step and hand-designed-based reconstructed medical image results are very blurry and introduce new noises in the images. Due to this, sometimes medical practitioners make wrong decisions, which are very dangerous for human beings. In this chapter, the authors explain that the hand-designed as well as deep learning-based approaches, including some image quality assessment metrics to open the gate to verify the images with different approaches, depend on the single image approach. Furthermore, they discuss some important types of medical images and their properties.
TopIntroduction
Reconstructing the visually pleasing high-resolution (HR) images from the low-quality or low-resolution ones is a challenging classical problem in image enhancement. This restoration process is called an image super-resolution (SR), which depends on pre-or post-processing steps to boost the perceptual quality of the recovered output image. In the field of medical images, upscaling the LR image into the desired HR image is no easy task due to some physical limitations of imaging systems and noise factor, jagged ringing artifact, and blurring output. Hand-designed filter approaches have already resolved these issues, but performances are not satisfactory. Recently, Artificial Intelligence (AI) has been involved in medical imaging and reconstructs the visually pleasing quality of the MRI image compared to earlier approaches. However, AI is a very new idea, which was begun in the 1940s. The name of artificial intelligence was invented in 1956 by John McCarthy. In other words, AI is used for computer algorithms that can mimic the human cognitive level. The most recent success of AI has been enabled by massive increases in both computer power and data availability. AI implementation in medical imaging improves the quality of clinical practice and supports the clinical decision (Hosny, Parmar, Quackenbush, Schwartz, & Aerts, 2018; Vickers, 2017; C. Wang, Zhu, Hong, & Zheng, 2019). Currently, machine learning technologies have matured to the point where they can meet clinical criteria. The development in information and data has been linked to improved illness knowledge and comprehension, thanks in part to advances in technologies that induce quantitative and qualitative measures of physiological labels. Indeed, the promise of machine learning as a platform for combining data from various references into an intertwined system may greatly benefit highly competent worker’s decision-making processes. When machine learning was still in its infancy, it was hypothesized that the capability to gather and store massive quantities of information in a knowledge base would determine the success of an intelligent system that could learn and improve (Clancey & Shortliffe, 1984). Deep learning (DL) is an area of AI that has exploded more popular in recent decades. Because of its adaptability, high performance, strong generalization capability, and diverse applications. A significant volume of medical data and the development of increasingly powerful computers have sparked interest in medical imaging enhancement.
A Deep Convolutional Neural Network (DCNN) is a type of CNN neural network type approach that could enlarge the input image to awesome gadgets withinside the image and distinguish one from the different. In evaluating different types of methods, CNN calls for considerably fewer pre-processing steps. The structure of a CNN is stimulated via way of means of the organization of the Visual Cortex, which is connected to the sample of neurons in withinside the brain. Different neurons respond to stimuli best in the Receptive Field, a tiny area of the field of view. Several comparable fields may be stacked on the pinnacle of every different from spanning the entire field of vision. Recently, with the advancement of image enhancement technology, especially in a medical image, SR has performed valuable clinical applications. MRI image gives a rich, detailed information of high resolution in medical images and their applications (Greenspan, 2009; Van Ouwerk & Computing, 2006). The main function of MRI is to obtain the sagittal plane, transverse plane, different inclined plane, and coronal plane (Huang, Shao, & Frangi, 2019) information. The medical images enhancement in terms of (resolution) is limited by different enlargement parameters, like scanning type hardware up-gradation, signal-to-noise ratio (SNR), moving objects, limited scanning time of hardware, technical and economic factors (J. Zhu, Yang, & Lio, 2019;), and image motility of organs. These factors decrease the perceptual quality of the medical images and generate very blurred images, having jagged ringing artifacts and diminishing the visibility of important pathological details. The primary goal of image SR is to reconstruct the highly pleasing quality of output MRI images from the degraded version of quality input MRI images, which is the typically ill-posed problem (X. Wang et al., 2018).