The late physicist Stephen Hawking was one of the most brilliant minds of his time. He believed that Artificial Intelligence (AI) could serve the good of the world. He also got the world’s attention when he warned about the danger that AI could eventually outsmart humanity. He urged that its creators exercise responsibility with best practice and effective management.

AI is a reality that touches all of us throughout our day. Common examples include easier rush hour commutes; sorting incoming email as useful (e.g. primary, social, promotion) while filtering out SPAM; connecting you to others through social media; and much, much more.

AI is the process of creating machines with their own intelligence. One component of AI is Machine Learning (ML) which includes specific scientific statistical methods that enable computers to learn on their own, progressively building on their “knowledge” without having to be repeatedly programmed by humans. In turn, a component of ML is Deep Learning (DL) which uses biological models like the brain’s neuron networks that operate bodily systems. With DL, computers can literally teach themselves to accomplish “human-like tasks, such as recognizing speech, identifying images or making predictions.”[i]

Artificial intelligence and radiology

With the advent of self-driving cars or airplane autopilots, we entrust our lives to the minds of those who create AI. A medical area in which AI is responsible for human life is radiology. This field began with medical imaging (X-rays, ultrasound, CT and PET/CT scans, MRI) to detect and diagnose abnormalities. The ability to “see” inside the body led to minimal-to-noninvasive image-guided procedures (interventional radiology) as an alternative to major surgery.

AI offers tremendous potential within radiology to improve patient outcomes. Thanks to the digitizing of diagnostic imaging, literally millions of images from all over the globe are available for computerized recognition and classification in milliseconds. AI systems, both ML and DL, are already in place to improve diagnostic accuracy and efficiency, and help radiologists make procedural decisions. With AI, a correct diagnosis based on advanced imaging may mean that correct treatment can begin the same day.

ML and DL techniques that support radiology diagnosis and decision-making

There are many techniques by which ML and DL enhance radiologic practice. Here are three types:

1. Rule-based reasoning takes advantage of human expertise to develop “if-then” logic. For instance, let’s say there is software programmed with knowledge about clinical bone problems. When the radiologist presents it with specific information and images from a patient’s case, the software applies appropriate rules to make an educated inference: IF imaging reveals a porous bone area, THEN the problem may be either osteoporosis or bone cancer.” Of course, this is a simplistic example. In actual clinical use, very sophisticated software is presented with complex, ambiguous cases and applies the rules in nanoseconds, helping a radiologist identify a condition and rule out other possibilities. 
2. Artificial neural networks are modeled on how the brain’s nerve cells (neurons) are structured into communication and learning networks. They use a large number of interconnected elements with statistically weighted “decision trees” to learn directly from observations. These can be used for perceptual tasks such as identifying tumor patterns in, say, a CT scan of the liver. (See my blog on computer diagnosis of prostate cancer.)
3. Hypertext and hypermedia allow nonsequential access of related materials from the “cloud.” If you clicked on the above link to my blog, you “hopped off” of this page to a related page that was out of this sequence, and then hopped back to pick up where you left off. Thanks to AI, a radiologist facing an ambiguous clinical case may not be certain if it’s condition A or B. With a simple “Tell me everything about condition A” command, a wealth of pertinent articles, radiologic images, charts, graphs, etc. can be quickly accessed and filtered to help the radiologist determine that it is or is not condition A and move on from there.

Developer responsibility

There are other ML/DL types as well, but I hope the above categories give you some idea of the kinds of software that are helping radiologists – and there are more in development every day. With human life and well-being is on the line, the burden falls to software programmers make AI safe by designing ML and DL using best practice and effective management. Over 20 years ago, a visionary radiologist from the University of Pennsylvania, Charles E. Kahn, Jr. foresaw the task ahead:

Builders of expert systems must choose the most appropriate artificial intelligence techniques for a particular application… In addition, developers must address organizational and operational aspects of a decision support system… Validation and evaluation are crucial. Validation assures that a medical decision support system’s advice is ‘accurate, complete, and consistent’; evaluation addresses the system’s applicability, speed, acceptability, and utility to physicians in clinical practice… Developers must establish an ongoing means to monitor and update a decision support system’s knowledge to prevent its gradual obsolescence.[ii]

While AI can never replace the personal doctor-patient relationship, it can equip the radiologist in a way that allows greater authority and expertise than the individual alone can have. It can accelerate information gathering and decision-making. It can increase the radiologist’s confidence in diagnosis and treatment planning. In short, responsible AI enlarges the mind of each radiologist with the brilliance of experts for the good of each patient.

---

[i] https://research.googleblog.com/2018/04/an-augmented-reality-microscope.html

[ii] Kahn, Jr, Charles. (1997). Artificial Intelligence in Radiology: Decision Support Systems. https://www.researchgate.net/publication/2665196_Artificial_Intelligence_in_Radiology_Decision_Support_Systems?enrichId=rgreq-c6b7508a7d86c6f1cb5c4345276df256-XXX&enrichSource=Y292ZXJQYWdlOzI2NjUxOTY7QVM6MjgzNzI4OTI5NjczMjE5QDE0NDQ2NTc2NTY5NTI%3D&el=1_x_3&_esc=publicationCoverPdf

[accessed July 18, 2018]