Deep Learning and Medical Applications

Digital medical images can be viewed as digital representations of real physical tissue properties that are convenient for handling, storing, transmitting, retrieving, and analyzing image information. In order to perform feature extraction/identification/classification from high-dimensional medical data, we need dimensionality reduction (DR), which can be achieved by identifying the statistical patterns in medical image data with highlighting their similarities and dissimilarities. Here, the dimension of medical images is the total number of pixels in the image. In DR, we try to find an useful representation of reduced dimensionality in high-dimensional data, which minimizes information loss by maximizing local data variance. Given medical image data, the key challenging issue is how can we efficiently extract a low-dimensional latent structure? Various dimension reduction techniques have been developed to process the high-dimensional data where the intrinsic dimensions are assumed to be much lower. In the very ideal case, data can be regressed linearly and DR can be performed by principal component analysis. This chapter explains the theories, principles, and practices of DR techniques.

Segmentation of a target object in the form of closed curves has many potential applications in medical imaging because it provides quantitative information related to its size and shape. Over the last few decades, many innovative methods of performing segmentation have been proposed, and these segmentation techniques are based on the basic recipes using thresholding and edge-based detection. Segmentation and classification in medical imaging are in fact experiencing a paradigm shift due to a marked and rapid advance in deep learning (DL) techniques. DL methods have nonlinear representability to extract and utilize global spatial features and local spatial features simultaneously, showing amazing overall performance in medical image segmentation. DL methods mostly lack transparency due to the black-box output, so clinicians cannot trace the output back to present the causal relationship of the output diagnosis. Therefore, in order to safely utilize DL algorithms in the medical field, it is desirable to design the models to transparently explain the reason for making the output diagnosis rather than a black-box. For explainable DL, a systematic study is needed to rigorously analyze which input characteristics affect the output of the network. Despite the lack of rigorous analysis in DL, recent rapid advances indicate that DL algorithms will improve their performance as training data and experience accumulate over time.

This chapter reviews metal artifact reduction (MAR) methods for low-dose cone-beam computed tomography (CBCT). MAR is of vital significance because the number of aged people with artificial prostheses and metallic implants is swiftly increasing with the rapidly aging population. Metallic objects present in the CBCT field of view produce streaking artifacts that highly degrade the reconstructed CT images, resulting in a loss of information on the teeth and other anatomical structures. Metallic object-related artifacts are associated with beam hardening, scattering, partial volume effects, and a high degree of inhomogeneous attenuation to name a few. As metal-induced artifacts are complex and nonlinearly intertwined, MAR has remained a challenging problem over the last four decades. Metal artifacts are caused mainly due to a mismatch in the forward model of the filtered back-projection (FBP) algorithm. The presence of metallic objects in an imaging subject violates the model’s assumption that the CT sinogram data is equal to the Radon transform of an image. FBP ignores the polychromatic nature of the X-ray data \(\textbf{P}\), which has nonlinear dependence on the distribution of the metallic object. Various MAR methods have been suggested, but the existing MAR methods do not reduce the metal artifacts effectively in low-dose CBCT environments and may introduce new streaking artifacts that did not previously exist. We hope that this chapter will help develop new MAR algorithms that overcome the limitations of existing MAR methods and effectively reduce metal artifacts to facilitate diagnosis, preoperative and presurgical assessments, surgical navigation, and workflows for rapid prototyping.

Digital dentistry is evolving rapidly with the rapid innovation of artificial intelligence (AI) and the advancement of an AI-based digital platform that integrates 3D jaw–teeth–face data from various imaging devices such as cone-beam computerized tomography (CBCT), oral scanner, face scanner, 3D tracking devices, and others. Digital dentistry equipped with the AI-based integrated platform enables dentists to provide accurate diagnoses and treatment while saving time and cost, significantly improving digital workflow. Additionally, digital dentistry also improves patient satisfaction by increasing the patient’s comfort and decreasing the chance of revisiting the dental clinic thanks to the enhanced accuracy level of dental treatment on the teeth, gingiva and occlusion. In digital dentistry, the 3D digital composite model obtained from image data of CBCT, and oral and facial scan will be an essential tool for almost all processes, including virtual treatment planning and on-screen simulation of surgical or dental treatment. Noting that the dental regions of 3D CT data do not have the level of resolution to be used directly for treatment, the jaw–tooth composite model, which accurately fuses the individual tooth geometry obtained from the dental impression model or oral scan with the jaw bone obtained from CT, is important for planning and performing dental treatment and predicting treatment outcome.

Diagnostic ultrasound is the most commonly used imaging method in the field of obstetrics and gynecology to estimate various biometrics related to fetal development, fetal well-being, and perinatal prognosis. Until now, ultrasound measurements of fetal health parameters (i.e., amniotic fluid volume, biparietal diameter, head circumference, abdominal circumference, and others) have been made through a cumbersome and time-consuming manual process, and their accuracy depends heavily on the operator’s skill and experience. Therefore, there has been a high demand for an easy-to-use interface for collecting biometrics from fetal ultrasound images to improve clinician workflow efficiency. Traditional methods have fundamental limitations in automating biometric measurements from noisy ultrasound images that are often degraded by signal dropouts, reverberation artifacts, missing boundaries, attenuation, shadows, speckles, and so on. Medical imaging is experiencing a paradigm shift due to the remarkable and rapid advancement of deep learning technology, and ultrasound companies, including Samsung Medison, are making every effort to develop a new AI-based system for automated fetal ultrasound diagnosis. The reason for these efforts of ultrasound companies is that AI technology is expected to become a turning point in diagnostic ultrasound. This chapter focuses on fetal ultrasound, explains deep learning-based medical imaging technology, and hopes to help readers discover new possibilities and to provide future directions.

Recently, there has been marked progress in electrical impedance imaging in which cross-sectional image reconstructions inside the human body are pursued. These techniques also have wider applications as imaging methods in medicine, biotechnology, non-destructive testing, the monitoring of industrial processes, and in other areas. Their imaging techniques allow to visualize new contrast information of biological tissues and organs exhibiting distinct electrical properties depending on their physiological functions and pathological states. The mathematical models for bioimpedance imaging are expressed as nonlinear inverse problems involving time-harmonic Maxwell’s equations with electrical tissue properties being described by frequency-dependent conductivity and permittivity. This chapter reviews electrical tissue property imaging modalities.

Recently, with the significant developments in deep learning (DL) techniques, solving underdetermined inverse problems has become one of the major concerns in the medical imaging domain, where underdetermined problems are motivated by the willingness to provide high-resolution medical images with as little data as possible, by optimizing data collection in terms of minimal acquisition time, cost-effectiveness, and low invasiveness. DL methods appear to have a strong capability to explore the prior information of the expected images via training data, which allows one to deal with the uncertainty of solutions to ill-posed inverse problems. This chapter aims to discuss some mathematical interpretations of DL-based nonlinear low-dimensional representations of expected solutions to ill-posed inverse problems.