Machines That Learn: Can They Learn to Interpret Radiographs?

TLDR: In this article, the authors present a rule-based artificial neural network (RNN) language for the interpretation of medical images, which can be used to obtain expert advice on various topics from car mechanics to strategic business planning.

Abstract: T hroughout history, humans have used tools to automate labor-intensive tasks. Employing machines to replace manual labor is commonplace, but automation of cognitive jobs is now also becoming possible. Generally, some form of artificial intelligence is required for a machine to assume cognitive functions. To this end, many artificial intelligence computer languages have been developed that allow users to obtain expert advice on various topics from car mechanics to strategic business planning. These are primarily rule-based languages that evaluate data in a predictable and orderly fashion. These languages have served well for problems in which the parameters and their interrelationships are completely known. Unfortunately, rule-based languages do not lend themselves to tasks, such as interpretation of medical images, that require the integration of visual data with medical knowledge in the absence of explicitly defmed rules. As a result, attempts to apply rule-based artificial intelligence techniques to radiologic image interpretation has generally met with clinically ineffective results [1]. Another type of artificial intelligence, known as artificial neural networks, shows promise toward handling cognitive functions such as pattern recognition and can be applied to medicine and to radiology in particular [2]. Artificial neural networks are so named because they are inspired by, but not necessarily modeled after, biologic neural systems, with which they show some uncanny similarities in behavior. Standard linear computers, present on many desktops, are based on the Turing model. Turing was an early 20th century British mathematician who theorized that all cognitive behavior could be simulated with a ticker-tape with holes punched in sequence to show ones and zeroes. Assuming that the holes are in known positions on the tape in a predefined code, a machine with a counter (or central processing unit) could store and analyze all types of informarion. Turing’s idea appealed to cognitive behaviorists of the day who believed that this machine for information storage and processing represented the way the mind functions. Although Turing’s machine, with a central processing unit and address-relocatable memory, has served as an admirable paradigm for modem computers, no evidence exists that human brains actually function in this manncr. Indeed, a large body of evidence exists to the contrary. For example, human brains do not appear to have any single central processing unit. Tracing cognitive function has been exceedingly difficult, and no investigator to my knowledge has ever been able to locate a single set of neurons that correspond to a discrete memory or set of memories from a particular period of life. In fact, brain ablation studies suggest that human brains store information diffusely through networks that may even contain other, unrelated information. The difference in design between linear computers and artificial neural networks is manifested in the type of problems each is best suited to handle. Linear computers are fabubus for routine numerical manipulations such as addition, subtraction, multiplication, and division and for logic problems (Fig. 1). Networks, on the other hand, are not. Their entire construction is designed for pattern recognition via integration of information. This is much like a human brain. Although most of us are not proficient at manipulating even moderate numbers in our heads, we are capable of recognizing patterns quickly (Fig. 2). In addition, although linear computers must handle information one piece at a time, neural networks-just like humans-integrate information. When posed with a complex decision, such as whether to buy the red or the black sports car, humans do not individually scrutinize each factor involved in making the dccision but rather integrate the information and reach a decision without conscious processing. An example may help to explain how information may be processed and stored in a neural network. Consider 1000 students in a study hall, each with a battery, switch, and ammeter that is calibrated to an arbitrary scale ranging from one to 10. Each student’s