AI: An Intelligent Agent
The foundation for this book about lung CT AI is the application of what Alan Turing described in 1936 as the “universal Turing machine.” This is what is known today as the computer hardware and software that dominates so much of our lives, and is at the heart of lung CT AI. In his recent book, Stuart Russell describes succinctly what Alan Turing meant by the universal Turing machine. The essence of Turing’s discovery was to define two new mathematical objects. The first was defined as a machine, what is known today as computer hardware. The second, mathematical object, Turing defined as a program that is known as software code that runs on the computer hardware. Together, the machine (computer hardware) and the program (software code) define a sequence of events, or a sequence of state changes, that occur in the machine (computer CPU, computer memory, etc.) to accomplish a task.
Russell describes the key concept in modern artificial intelligence (AI) as being the concept of an intelligent agent. The intelligent agent exists in the software programs running on a computer. How the AI agent is built depends on the objective(s) to be achieved or the problem(s) to be solved. The functioning AI agent then depends on four important things: (1) environment; (2) observations; (3) actions; and (4) objective(s) ( Fig. 1.1 ). The environment is the physical and electronic space that the AI agent can access. Using the word processing (WP) AI agent as an example, the WP AI agent environment includes keyboard commands, computer display, and computer hardware and software, as well as any internet connections that are running. The observations that the WP can make are the keystrokes pressed, and it can read the WP files that exist on the computer and in the cloud. The actions that the WP can take are: recording the keystrokes and displaying those on the screen; storing them on the computer or the Internet; and reading and writing existing WP files from the computer memory, hard disk drive, and the Internet ( Fig. 1.2 ). The objectives for the WP AI agent are determined by the people who wrote the software to run on the computer. These objectives can be summarized as taking keystroke inputs and creating a software file that records the keystrokes and displays them on the computer screen. It is also important to recognize that one AI agent can pass the objectives to another AI agent to perform additional objectives, and this process can continue with as many AI agents as desired. For example, the WP files on a computer can be sent to a typesetting program that a publisher would use to generate the final output for a book.
AI Definitions and Levels
AI is a rapidly expanding field, and definitions of the different levels of AI are changing as a result of this growth. AI, for the purpose of this book, includes four levels: (1) reactive machine; (2) limited-memory; (3) theory of mind; and (4) self-aware ( Box 1.1 ). Since the late 1960s and early 1970s, reactive machine type AI has driven the development of multiple AI technologies to visually and, subsequently, quantitatively assess the presence and extent of lung diseases using x-ray CT scanning. More recently, limited-memory levels of AI have been added to the list of technologies driving progressive improvements in the visual and quantitative CT assessment of lung disease. Reactive machines are the most basic type of AI system. This means that they cannot form memories or use past experiences to influence presently made decisions; they can only react to currently existing situations—hence the term “reactive.” An existing form of a reactive machine is Deep Blue, a chess-playing supercomputer created by IBM in the mid-1980s. Reactive machine AI programs will only react in the present in the way they are programmed. Examples of reactive machine AI programs in CT lung AI would include analytic CT image reconstruction algorithms, analytic CT image lung segmentation programs, computing the lung CT image voxel histogram, and relevant voxel histogram statistics that have been previously shown to correlate with normal and diseased lung tissue. These reactive machine lung CT AI metrics of lung disease can be used to detect and assess the extent of normal and abnormal lung structure and function caused by underlying lung disease, such as emphysema, asthma, or pulmonary fibrosis. The advantage of reactive machine learning AI is that it is very clear what the software program is doing. However, it is not as powerful as limited-memory AI. Limited-memory AI is comprised of machine learning models that derive knowledge from previously learned information, stored data, or events. Unlike reactive machines, limited-memory learns from the past by observing actions, or data fed to them, in order to build experiential knowledge. Limited-memory machine learning AI has been used in identifying specific patterns of lung disease, such as honeycombing in patients with ILD. Limited-memory machine learning AI has also been applied to CT image reconstruction, reducing noise in reconstructed CT images. It has also been implemented in segmentation software to extract the lungs from the rest of the thoracic anatomy on CT images. Limited-memory AI machine learning has been used to identify unique patient CT phenotypes in patients with COPD and asthma. Limited-memory AI machine learning has also been used to classify lung nodules into benign and malignant categories.
Lung CT AI involves a number of simpler AI agents that are linked together to produce a final lung CT AI objective, which is to detect and assess normal and diseased lung structure and function and make these results widely available to patients and healthcare providers. The lung CT AI agents that are included in this book, in the order in which they are usually performed, are: (1) generate high-quality CT images of the thorax; (2) display the CT images of the thorax; (3) separate or segment the lung CT images from the rest of the thoracic anatomy; (4) extract quantitative features from the lung CT images that represent meaningful metrics of normal and diseased lung structure and function; (5) analyze these features to predict the presence or absence of lung disease, and to assess the extent and impact of any detected lung disease; and (6) make these results available in near real time for patients and their health care providers ( Fig. 1.3 ). The routine lung CT AI outputs for individual patient care can then be systematically collected and examined across healthcare systems and countries; these results can be used in near real time to assess world lung health and disease, and to inform how best to allocate scarce resources needed to decrease the prevalence and severity of lung disease. My biggest reason for writing this book is to make everyone aware that the lung CT AI objectives outlined above are now achievable. I believe lung CT AI will work best in a healthcare environment where everyone has access to quality healthcare, regardless of ability to pay or preexisting conditions.
This book will follow the sequence of AI agents that have been developed over the last 50 years to generate a 3D digital representation of the lung using x-ray CT, and finally develop an AI agent that is useful in detecting and predicting the malignant potential of pulmonary nodules; the presence and severity of diffuse lung diseases, such as chronic obstructive pulmonary disease (COPD), idiopathic pulmonary fibrosis (IPF), and, most recently, COVID-19 viral pneumonia. The progressive improvements in CT scanner technology have advanced the 3D visualization of normal lung structure and function and made it possible to develop new AI methods to accurately detect and assess normal and diseased states of the lung. The progressive increases in our knowledge of lung CT AI will help guide our narrative from describing the first digital images of the lung obtained from x-ray computed tomography, to the very recent exciting application of quantitative CT AI methods for the detection and assessment of the severity of lung nodules and acute and chronic diffuse lung disease. The high spatial and contrast resolution of modern chest CT images of the lung has also enabled the creation of sophisticated software to build silicon computer models of normal and diseased lungs, and has increased our fundamental understanding of lung physiology and pathophysiology.
Diagnosis of COPD, ILD, Lung Cancer, and Other Smoking-Related Diseases
The inhalation of cigarette smoke or other environmental combustion products is the leading cause of COPD and lung cancer. The inhalation of these combustion products leads to inflammation and oxidative stress in the lung tissues. COPD decreases the effectiveness of gas exchange through the destruction of alveolar walls and increases the resistance of getting gas to the alveoli through the narrowing and destruction of airways. The combination of these processes greatly impairs the function of the lungs in a COPD patient. COPD is the fourth leading cause of death in the United States behind heart disease, cancer, and accidental deaths. COPD is usually diagnosed late in the course of the illness and is often misdiagnosed. Lung cancer screening in the United States, using low-dose chest CT scans, is an opportunity to detect early lung cancers caused by exposure to cigarette smoking, but also to diagnose COPD at earlier stages when there is more time to intervene and alter the course of the disease.
Earlier detection of chronic interstitial lung disease (ILD) provides the opportunity for early therapeutic intervention and improved patient outcomes. Chronic ILD produced by IPF, HP, and CTD produces thickening of the alveolar walls, which decreases the efficiency of gas exchange. These diseases can also distort and destroy the normal architecture of the lung acinus, decreasing the effective surface–to–volume ratio of the lung. ILD is also a disease that is often diagnosed in more advanced stages and often misdiagnosed. People with IPF frequently have a significant history of cigarette smoking, and ILD is often present in patients with COPD.
Information for Healthcare Providers and Administrators, Patients, and Researchers
I believe there is a need for a book on lung CT AI to inform healthcare providers and administrators, patients, researchers, and government agencies about the development, validation, and commercial availability of lung CT AI products that can detect and assess several lung diseases, including lung cancer, COPD, COVID-19 pneumonia, and ILD. COPD and ILD are often diagnosed in later stages of the disease and often missed in the earlier stages of disease when the disease is more treatable. The lung CT AI assessment of COVID-19 viral pneumonia during the ongoing worldwide pandemic of 2020 has been helpful in countries and healthcare systems where access to timely quality RT-PCR testing is not available or is unreliable. Lung CT AI programs can help identify COVID-19 pneumonia from other forms of pneumonia and assess which patients will need to be hospitalized and are at increased risk of dying.
Lung cancer is the leading cause of cancer deaths in the United States, and there now exists AI software that can help identify benign versus malignant lung nodules. Until recently, these lung CT AI technologies were only used in research studies because of the difficulty in deploying the software in busy radiology practices within clinics and hospitals. This has changed with the advent of software programs, like VIDA Insight, that can be deployed independently or in larger enterprise AI ecosystems of large computer companies. Small agile software companies, like VIDA, can develop specialized AI software to assess lung CT images for the presence and severity of lung structural and functional changes due to disease, and these specialized AI software programs can now be accessed at the point of care without slowing down the radiology workflow; this novel CT lung AI software can increase efficiency in radiology practices. There are potential broader favorable impacts on human health than just the care of individual patients. QCT metrics of lung disease that are driven by environmental pollution, such as COPD and lung cancer, can be assessed in an objective way across hospital systems, states, and nations. This will inform governments and leaders on how best to spend limited resources on improving the environment and lessening the spread of environmentally driven lung diseases.
Describing Lung CT AI in Three Stages
This book includes a historical description of the technological developments that were necessary to achieve the current success of lung CT AI software in the clinical care of patients with lung disease. The structure of the book is divided into three segments. The first segment, Chapter 2, Chapter 3 , discusses the development of x-ray CT scanners and scanning protocols that are used to scan the thorax and generate 3D images of the lungs.
Chapter 2 begins with the invention of the x-ray computed tomographic (CT) scanner that, at first, could only generate a limited number of low-spatial-resolution contiguous 2D digital images of the thorax with a long scan time. The first whole-body CT scanner, ACTA CT scanner, had very slow scan times, 4.5 minutes to obtain two 7.5-mm nearly contiguous axial images of the thorax with a spatial resolution in the axial x-y plane of 1.5 mm and the z-axis of 7.5 mm. The ACTA CT scan time for an adult male lung that is 30 cm in length would take a minimum of 90 minutes, longer if x-ray tube cooling was needed. The scan time and the spatial resolution of the CT scanner are important metrics that drive what can be done with AI in analyzing images of the lung. Chapter 2 also follows several key technologies that were developed to improve the spatial resolution and decrease scan times of x-ray CT scanners; scan time and spatial resolution have greatly improved over the last 40 years.
Chapter 3 discusses the latest generation of x-ray CT scanner technologies and lung CT scanning protocols available in 2020. The latest generation of CT scanners can scan the entire thorax in less than 10 seconds with an isotropic resolution of 0.5 mm. Chapter 3 details the important CT scanning variables that need to be carefully selected, such as x-ray dose, scan time, z-axis resolution, and image reconstruction method, to produce the best lung CT images.
The second segment of the book, Chapter 4, Chapter 5, Chapter 6, Chapter 7 , describes the lung CT AI methods that have been developed to detect and quantitatively assess focal lung nodules, pulmonary emphysema, pulmonary fibrosis, and acute COVID-19 viral pneumonia from the CT-generated density maps of the lung.
Chapter 4 discusses the concept of the lung nodule and how CT is used to detect and quantitatively assess the malignant potential of a lung nodule. Exposure to environmental factors, especially cigarette smoke, increases the risk of developing lung cancer. It has been shown recently that the use of screening lung CT scans can decrease mortality in people exposed to cigarette smoke.
Chapter 5, Chapter 6, Chapter 7 introduce increasingly more sophisticated lung CT AI methods to assess diffuse lung disease, starting with CT image voxel histograms in Chapter 5 and ending with sophisticated limited-memory AI machine learning algorithms in Chapter 7 .
Chapter 5 describes the QCT metrics that are readily obtained from a single total lung capacity (TLC) CT scan done with an appropriate CT protocol, as outlined in Chapter 3 . Lung CT density metrics were the first lung CT AI metrics to be reported and focused mainly on the reduced density of the lung produced by pulmonary emphysema. Pulmonary fibrosis produces increased density in the lung and can also be assessed using density measures from lung CT images. Chapter 5 describes the research studies that were performed to determine if the QCT metrics described do, in fact, represent important features of normal and diseased lung tissue by correlating CT image findings to other measures of lung disease (e.g., pulmonary function tests, lung pathology). Chapter 5 discusses how quantitative CT lung images provide critical new information regarding COPD that was not obtained from other methods (e.g., clinical history, pulmonary function testing). Determining the value of lung CT AI versus other means of assessing patients with COPD required the funding of large multicenter NIH grants that studied thousands of subjects with and without COPD. These studies established the relative value of lung CT AI metrics with other data characterizing COPD-related lung disease, including genetic studies, pulmonary physiology testing, and standardized healthcare questionnaires. These studies included COPDGene, MESA Lung, and SPIROMICS. One QCT lung AI study in the SPIROMCS study identified four unique clusters of subjects with COPD that had different QCT metrics and different disease trajectories. Chapter 5 also discusses the results of multiple smaller-size research studies in assessing the value of lung CT AI in groups of patients with pulmonary fibrotic lung disease associated with idiopathic pulmonary fibrosis (IPF), connective tissue disease (CTD), and hypersensitivity pneumonitis (HP). A number of QCT density metrics are shown to be effective in assessing pulmonary fibrosis in ILD. These metrics include whole lung histogram measurements of the lung and their associated statistics, including mean lung density, skewness, and kurtosis. They also include QCT measures of the amount of lung density between -600 HU and -250 HU.
Chapter 6 introduces dynamic QCT metrics that can be generated using two CT scans done sequentially at different lung volumes. These metrics provide a means of assessing lung structure and function. The assessment of lung ventilation can be done using an inspiratory and expiratory CT scan and looking at the change in lung density, or change in lung volume, between the two scans. There are several ways to apply this methodology to assessing lung ventilation at the whole lung level, lobe level, segment level, acinar level, and voxel level. Lung biomechanics can also be assessed using two chest CT scans obtained at different lung volumes, typically TLC and RV.
Chapter 7 looks at limited-memory type AI metrics in assessing COPD, ILD, and COVID-19 pneumonia. Limited-memory type AI algorithms are trained in three stages ( Fig. 1.4 ) to detect and assess the presence of lung disease. The first stage is feature extraction. This can be done using supervised or unsupervised approaches. Supervised approaches have an expert imaging physician label the normal and abnormal tissue features on the lung CT images. Unsupervised approaches let the AI algorithm determine the normal and abnormal tissue types. The second stage used to train the AI algorithm is to have it recognize and quantitate the amount of the normal and abnormal tissue features, or textures, within the lung CT images that correspond to important pathologic features associated with different lung diseases. The number of supervised or unsupervised training CT cases can be increased to improve the performance of the limited-memory lung CT AI program. The design of the limited-memory lung CT AI program can also be altered to improve its performance. The third stage tests the performance of the trained AI algorithm on a test set of chest CT scans from a cohort of human subjects separate from the training cohort. The performance of the lung CT AI algorithm is based on its ability to identify the different tissue features on the test set of chest CT scans, including features such as emphysema, honeycombing, and consolidation.