By Cameron Simmons and Aaron Brantly
Doctor Lewis Zimmerman looks real, he treats patients, moves in a physical space, answers questions, and would pass the Turing and Lovelace tests, but he is an AI doctor on the science fiction show Star Trek Deep Space Nine. Yet as the show opens Benjamin Sisko, captain of the space station expresses doubts about having an artificial intelligence for a doctor. Doctor Zimmerman in the spirit of the show jokes with the reluctant Captain Sisko and eventually they are able to work out their differences and the show moves on introducing occasional plot twists that place the fate of the living occupants in the hands of an often emotional and moody AI. We are a long way from having AI doctors help with every condition, but in the field of Radiology, the future is now. Radiology is a subfield of medicine that utilizes imaging technology to diagnose and treat disease. Radiologists, trained medical doctors, are experts at identifying and diagnosing abnormalities. They then provide their diagnosis to other physicians who then take the next steps. Radiology is seemingly uniquely positioned for takeover by AI. Radiologists are required to know and understand disease from many different perspectives. As a result, well-trained radiologists must know what specialists in multiple different subfields of medicine are looking for. This requires radiologists to develop a large repository of information in which they create mental models that helps them to identify and quickly classify abnormalities. The radiologist’s job seemingly matches up to the description of pattern recognition systems that first began to arise in the 1960s and have become increasingly prevalent in everything from iPhones to self-driving cars.
An increasing volume of research is being undertaken to move radiologists to obsolescence in computer science and biomedical engineering programs around the world. At first glance the problem appears to be straight forward. Radiologists are engaged in pattern recognition. They leverage their medical skill to help them identify abnormal patterns using different forms of imaging systems. Radiological exams are expensive and time consuming. Often it takes radiologists 30 or more minutes to “read” a single study. Why not simply substitute pattern recognition AI systems for radiologists? Would this save money, improve efficiency, and perhaps more importantly improve or maintain the same level of care? The answer to these and other questions is surprisingly complicated. Although artificial intelligence has dramatically improved the imaging capabilities of Radiologists over the last 20-30 years with advances in Magnetic Resonance Imaging (MRI), Computed Tomography (CT), Digital X-Rays, Dual-energy X-ray absorptiometry scans, Ultrasound scans, and more the assistive nature of AI in radiology has not translated fully to diagnostic replacement.
Radiology is a subfield of medicine that utilizes imaging technology to diagnose and treat disease. As a result, radiology is one of the subfields of medicine that is perfect for artificial intelligence. First and foremost artificial intelligence excels at pattern matching and has been known to see abnormalities in imaging that can sometimes go unnoticed by trained professionals. One of the important factors of modern medicine is speed and the quicker something is detected the easier it is to be treated, the more the symptoms can be reduced and, in some cases, can be the difference between life and death for patients. So, if radiology is so good at these things why don’t we just stop training radiologists altogether and put our focus into artificial intelligence enabled machines. The answer to that question is difficult as it is as simple as it is complicated. A radiologist’s job does not simply rely on doing pattern matching, there are much more complicated aspects of their jobs that with the current level of technology cannot be replaced.
Where does the problem for AI reside and why can’t it replace radiologists? It is true that pattern recognition AIs are improving, but enormous problems remain, first most image recognition AIs rely on large volumes of training data. This data has led to the creation of training sets that then output probabilities of disease. But the decision as to the accuracy of that probability remains a human decision. Moreover various modes of learning data pattern recognition AIs have been known to fail disastrously, such as when Google’s image recognition software identified African American faces as gorillas. Worse still Google didn’t actually fix the problem, rather it removed Gorilla as a label. The present state of AI is stuck in what is referred to as narrow AI. This means that AI’s “learn” from prior data. But the term “learn” is a complex term that implies value. By contrast many AIs simply are trained to identify pixel characteristics that approximate the characteristics of “known” data. Because these algorithms are often iterative, they are also subject to introduced initial biases (specific images limit the scope of possible types of an image) and induced biases over time (images over time foster a sub-sector of diversity that is not representative of a whole). Yet despite the present state of AI and Radiology there are indications that the issue may be pressed forward in the coming years as medical students, fearful of being overtaken by robots are avoiding specializing in radiology. AI will impact radiology, but there are unlikely to be any Doctor Zimmermans any time soon. Rather it is likely that the radiologist will be augmented by AI. They will not be augmented in a creepy, chip under the skin way, but rather more like what has occurred over the last 20 years with improvements in technology. AIs will be additive to the field of radiology and enable efficiencies of their human operators. The job of radiologists will remain in part pattern recognition, but like all physicians, radiologists also play a human role in interacting with patients and physicians. They will still “consult with other physicians on diagnosis and treatment, treat diseases (for example providing local ablative therapies), perform image-guided medical interventions (interventional radiology), define the technical parameters of imaging examinations to be performed (tailored to the patient’s condition), relate findings from images to other medical records and test results, discuss procedures and results with patients, and many other activities”. In a perfect world AI systems will handle the repetitive tasks of pattern matching and image scanning and assist radiologists in abnormality identification thereby, lowering overall costs for treatment and improving the accuracy in treatment plans crafted by radiologists and other physicians.