The topic of artificial intelligence (AI) has been quite prevalent in the mainstream press recently, with several high-profile warnings about the “singularity”, or the point at which computers will become self-conscious and smarter than humans, thus posing an existential threat to mankind. The fascination with computers' potential ability to emulate human behavior can be traced back to the 1950s, when Alan Turing posed the famous question “Are there imaginable digital computers which would do well in the imitation game?”, and the “Turing test” is widely known as a yearly contest between humans and computer programs, with a panel of human judges striving to correctly classify them by engaging each in a 5-minute typed conversation. There are many practical, technological, strategic and even philosophical considerations associated with the Turing test, leading us to conclude that computers are currently far superior to humans in their ability to quickly sort through, collate, reproducibly measure and analyze large amounts of data, but not as good at understanding context. With that in mind, this talk will focus on the admittedly humbler task of describing some of the ways in which computers and AI techniques can help and are being utilized in medicine today, with specific examples pertaining to computer-aided interpretation of medical images.
The main distinguishing criterion in AI is between knowledge-based (rule-based) algorithms and machine learning approaches. To offer an analogy with human intelligence, AI can express itself by a) applying rules it has been taught by an expert teacher, or b) inferring those same rules by looking at a substantial number of examples given by the teacher. These two categories have been traditionally referred to as “expert systems” (ESs) and “artificial neural networks” (ANNs), respectively.
A typical AI application that can be either knowledge-based or machine learning-based is the development of clinical decision support systems (CDSSs) to reduce misdiagnosis and improve patient treatment. At his core, this is a “big data” application, in which a supercomputer (such as IBM's Watson of Jeopardy's fame) searches through a large database of past patients' symptoms, diagnoses, treatment histories and outcomes, as well as medical literature and medical societies' guidelines, to offer the consulting physician a recommendations on the diagnosis and choices of treatment for a current patient. A CDSS can be integrated with a hospital's electronic health record (EHR), and a supercomputer is not actually needed, since all information and algorithms can reside in the “cloud”. The current thinking is that CDSSs should only provide an informed “second opinion”, not only for legal reasons but also because the data they examine is “structured” to reduce complexity, so that, for example, image parameters are considered instead of the actual images of a patient's study. Despite its potential, cost effectiveness of CDSSs is still being debated.
A relatively straightforward medical AI application well suited for the machine learning approach is speech recognition, increasingly used in the context of digital dictation systems for report generation. Conversely, a natural language report can be generated automatically based on physician input as well as stored demographic and quantitative (from laboratory or imaging procedures) information on a specific patient, with the benefit of rule-based internal consistency checks for the final output.
A very common application of AI in medicine is computer-aided interpretation of medical images, usually based on pattern recognition. Anatomical organs of interest such as the brain, the lungs and the heart, or pathological structures such as cancerous nodules and polyps can be isolated (“segmented”) using a combination of pre-processing techniques (de-noising, resolution enhancement), standard mathematical operators (to detect edges, contiguity, etc) and prior knowledge about expected size, shape and location, as well as comparison to normal databases for patients having undergone similar imaging procedures with comparable imaging protocols. Segmentation is typically followed by automatic or semi-automatic quantification of key parameters characterizing anatomy, physiology and pathology, such as standardized uptake values of tumors or cardiac contraction dyssynchrony. Of note, AI-based quantitative assessment of images is vastly immune to intra- and inter-operator variability, having potentially perfect reproducibility if based on an entirely automated algorithm, and is therefore ideal for the serial assessment of patients undergoing therapy.
Disclosure of Interest None declared