Machine Learning for Tomographic Imaging. Professor Ge Wang

of deep learning techniques in the CT field.

      The third part has chapters 6 and 7 on magnetic resonance imaging (MRI), in parallel to chapters 4 and 5. Chapter 6 reviews the MRI data acquisition process and the MRI scanner instrumentation. Fourier transform and compressed sensing algorithms are first presented. Then, classic post-processing algorithms are discussed. Chapter 7 covers various deep-learning-based MRI techniques, including a variety of deep reconstruction networks with applications to regular MRI, parallel MRI, dynamic MRI, and magnetic resonance fingerprinting (MRF). Miscellaneous topics are also covered, such as optimal k-space sampling and activation functions for complex-valued inputs. Finally, we discuss the integration of MRI data acquisition and image reconstruction with a synergized pulsing and imaging network (SPIN).

      In the fourth part, we offer chapters 8–10. Chapter 8 briefly presents other imaging modalities including nuclear imaging, ultrasound imaging, and optical imaging in terms of working principles, and then describes representative neural networks developed for these imaging modalities individually. After that, we mention multi-modality imaging. Chapter 9 discusses image quality for general and task-specific assessment. In this chapter, network-based model observers are presented as a new approach for cost-effective reader studies. Chapter 10 is on quantum computing. We start with wave–particle duality and quantum puzzles, define quantum bits and gates, and touch upon quantum algorithms and quantum machine learning.

      For your convenience, the relationships among the four parts and the associated chapters are summarized in figure 0.7, supplemented by appendices A and B. It is underlined that appendix B and associated web resources are under development, and should be invaluable to enhance the learning experience and AI/ML skills. As shown by this book, AI/ML techniques are applicable and instrumental to all tomographic modalities, and promise to unify individual modalities computationally.

      Figure 0.7. Diagram suggesting the order in which the reader reads the components of this book.

      0.6 More to learn and what to expect next

      As implied by figures 0.5 and 0.6, there are too many relevant papers to read, and the number of such papers is growing rapidly. After reading this book systematically or selectively, you will need more time to master more materials, dive in more deeply, and practice for better skills. According to PwC (https://www.pwc.com/gx/en/issues/data-and-analytics/publications/artificial-intelligence-study.html in June 2017), AI will yield a global GDP increase of 14% in 2030, and a top area AI affects is healthcare. In such an expanding phase of AI/ML R&D, we have no choice but to pursue continuous learning—papers, books, online materials, and hands-on projects.

      Despite all the above positive comments on AI/ML, this field has previously experienced two winters, and it is natural to wonder if we will sooner or later enter another winter of AI/ML. All this depends on how much and how quickly we can continue advancing the field and meeting the majority’s expectations. Although the future is often unpredictable but sometimes indeed inventable (Virginia Tech’s old logo ‘Invent the Future’), we tend to be very optimistic about the future of AI/ML in the long run, and are particularly hopeful for several directions of development.

      There are two scientific approaches for reasoning—deduction and induction. Accordingly, we have two associated schools of AI/ML. Deduction goes from general to specific, and is a top-down approach. Decades ago, research on rule-based expert systems was popular, and the fifth generation computer was a hot topic. In this context, it was hoped to reason from general rules to specific claims. On the other hand, induction works from data toward knowledge or information. This is bottom-up or data-driven. The recent champions of the ImageNet contest developed their deep learning programs in a data-driven fashion. There are great opportunities to merge these two approaches in the future. Knowledge graphs and self-supervised learning are two ideas along this direction.

      Also, AI/ML and neuroscience/psychiatry are closely intertwined and mutually promoting. For example, studies on the human vision system are an integral part of neuroscience research, which played an instrumental role in the development of AI/ML and suggested a mechanism for cellular-level processing and interconnection for visual perception and image analysis. New findings and insights in AI/ML and neuroscience will continue promoting each other. Arguably, human intelligence is closely related to natural language understanding and expression. It is believed by many that an important direction of AI/ML development is natural language processing (NLP) (figure 0.8).

      Figure 0.8. NLP is believed to be an important direction for development of AI/ML.

      Yet another outside-the-box approach is quantum computing, but no-one is sure when its prime time will come. However, if it becomes practical, AI/ML will be revolutionized. For example, our proposed 'SPIN' network may be implemented via quantum computing. A few days ago, Google announced a quantum supremacy, and heated on-going discussions on this topic (https://www.sciencenews.org/article/google-quantum-computer-supremacy-claim). We cannot exclude the possibility that intelligence is essentially a quantum phenomenon and must be implemented through quantum computing. Let us continue making and enjoying our AI/ML related efforts.

      References

Wang G 2016 A perspective on deep imaging IEEE Access 4 8914–24

Wang G, Ye J C, Mueller K and Fessler J A 2018 Image reconstruction is a new frontier of machine learning IEEE Trans. Med. Imag. 37 1289–96

IOP Publishing

      IOP Publishing

      Machine Learning for Tomographic Imaging

      Ge Wang, Yi Zhang, Xiaojing Ye and Xuanqin Mou

      Chapter 1

      Background knowledge

      1.1 Imaging principles and a priori information

      1.1.1 Overview

      Tomography is an imaging technology that studies an object with externally measured data generated by some physical means such as x-ray radiation, where data are projections in the form of line integrals of an object function from different angles of view. This kind of imaging technique can be used to produce images of hidden 3D structures in an opaque object non-destructively, even though the object is not transparent to the human eye. Seeing through a patient’s body is highly valuable for medicine. Hence, modern medicine is, in a sense, enabled by tomography.

      With major discoveries in physics, such as x-ray radiation and magnetic resonance, an object can now be imaged using various mechanisms. X-ray photons easily penetrate most every-day materials, including human tissues, and produce line integral information on the linear attenuation coefficient that characterizes the interactions between the materials and x-ray radiation. Various types of materials have different linear attenuation coefficients. From sufficiently many x-ray projections, a cross-sectional image can be reconstructed, which is called computed tomography (CT). In such a reconstructed image of a patient, bone, soft tissue, and fat can be clearly discriminated, and anatomical features can be well defined. Magnetic resonance imaging (MRI) is another important imaging modality. It takes advantage of nuclear spins (in particular spins of hydrogen nuclei) to generate signals when the spins are aligned in an external magnetic field, excited by radio frequency pulses, and then relaxed to their steady states. During the relaxation process, various tissues have different proton densities and take different times to relax, thereby