Artificial Intelligence in Radiology with Matthew Stephens, MD
Episode 82 Recap of the FLATLINING Podcast
This episode is the next in the series from Flatlining featuring medical professionals of various backgrounds discussing a wide range of topics. Dr. Matthew Stephens, a board-certified Radiologist who works at Radiology of Indiana holds a master’s degree in computer science and completed a fellowship in Medical Informatics at the Regenstrief Institute of Indianapolis. He joins Matthew and Ron to discuss the integration of Artificial Intelligence in the practice of Radiology.
Dr. Stephens related that he came to the practice of medicine in a roundabout way, the son of an Oncologist in Muncie Indiana, credits his father’s thirty years of working with patients that inspired his interest in medicine and science in general. During his undergrad work in pre-med at Indiana University, he took some computer science classes and decided to pursue a master’s in computer science. Following completion of his advanced degree, he went to work for the Center for Medical Genomics at Indiana University as a programmer. After three years of work there he entered medical school. He said that decision was based on his desire to work with patients and have that human interaction that he found lacking in his work as a programmer. He explained that it was his knack for pattern recognition and a talent for working visually that he said drew him to radiology.
Matthew asked how computer science and health informatics are impacting the practice of medicine. Dr. Stephens related the biggest impact is the pace at which medicine is practiced today. He said that compared to how his father was practicing when paper charts were the norm and doctors dictated their notes made the overall process was slower compared to today. He explained that medical informatics has standardized the handling of information. Where in the past information may not reach a referring physician for weeks or months, today information is exchanged at a much quicker pace. He said not only are doctors seeing more patients and at a higher rate, but patients are also getting test results much faster. He explained none of this was possible twenty or thirty years ago before all the standardization into an organized workflow from one hospital system.
Moving the discussion to the topic of artificial intelligence (AI), Matthew asked Dr. Stephens how it is affecting radiology. Dr. Stephens shared that he studied AI while earning his master’s in computer science and explained that AI was heavily used in the study of genomics research decades ago. According to Dr. Stephens, the ideas behind AI are not new, that is, automating computers to find patterns in data. He said what is different now is “They have figured out how to take some of these models that already existed, like neural networks, and learned how to make them more efficient.” He said basically in theory these neural networks can learn any pattern you give it if you have the right data.
What was holding AI back he said was the “computational expense.” He explained that if the pattern you wanted the neural network to learn was very complex, the bigger the model needed to be increasing the processing requirements. Hardware designs evolved and improved the neural networks of processors from operating them in series to stacking them in parallel, creating “deep neural networks.” Which gave them the ability to work through complex algorithms more quickly and efficiently. All these advancements sped up the training of the pattern recognition algorithms.
Dr. Stephens took us on a deep dive into the computer science of how improved technology has affected the practice of radiology. He described what are called “convolution neural networks” which in its simplest terms use complex computations to increase the accuracy of pattern recognition. This type of breakthrough allowed for much better image recognition. He said before the introduction of convolution neural networks the accuracy of pattern recognition was somewhere around 70% and now with this improved technology well above 95% with some types of diagnosis as high as 99%.
This is all very exciting for radiologists, he said since they deal with a lot of images, and these often have common or routine patterns the physician recognizes right away. He said much of the work for radiologists can be tedious and repetitive, the technology has the potential to automate some of their work, mainly those items that are repetitive but still need to be characterized and put in reports because they could be clinically significant even though most are not. The hope, he said, is that by automating some processes, the radiologists can read more images that algorithms are not very good at detecting. He said he thinks that their workflow will gradually change over time with this improved technology.
Matthew asked Dr. Stephens if there were images AI was better at recognizing than others. Dr. Stephens explained that it comes down to how common is an image, and how much variability the pattern you are looking for has. Something that has little variation and is very common, that type of image the technology would be very good at spotting, he listed a few examples, like the aorta or a pulmonary nodule, or a broken bone, something that has a very distinct appearance. Things that have higher variation, are tougher for the technology to recognize. He used the example of screenings for breast cancer which, he said, is highly variable in appearance. Pointing out that it would still require a physician to review. He added that just the nature of breast scans and chest X-rays with the images superimposed can be challenging for the technology, but it is also challenging for the physician. The goal he said is to improve the technology so it is at least as good as the physician.
Ron acknowledged the usefulness of AI as a tool and asked Dr. Stephens if AI could replace the radiologist. Dr. Stephens pointed out that as long as you have someone who is trained and understands the limitations of the technology AI will be a useful tool in the workflow. He said you could not turn AI loose and hope it all works out because models change over time and scanning technology improves and you have to retrain the models with these advances. He pointed out that AI will always require supervision. He envisions that the technology could do the initial screening and place those scans that could be questionable in the cue for review from the radiologist. Then human physicians will make sure the AI recommendation is accurate. This will go a long way to improving efficiencies for radiologists, he said.
Dr. Stephens conceded that even at a 99% accuracy rate when extrapolated over a million exams, are still a lot of misses, but he pointed out, that the human radiologist is not going to catch everything either. He said AI has the potential to make radiology more accurate in some respects and more efficient. He said, “It’s still up to the radiologist to have the effort to go through the study and look at it, you can’t just totally rely on the algorithm to do your work.”
Dr. Stephens does think that some parts of reading scans could become somewhat automated because the algorithms will be that good. He used the area of mammography that may get to the point where the technology matches or exceeds the human reading of scans, but as scanning technology or AI technology improves there will always be the striving for better diagnosis.
Ron related the challenge of physician burnout within radiology and asked if this technology could help. Dr. Stephens agreed that AI could make radiology and their processes more efficient, which could in some respects increase the ability to read more exams. However, he placed the idea of physician burnout among radiologists as more of a cultural issue. Saying that burnout is more about working people to the breaking point. Improvements like more breaks, and more social interaction are key factors that could help in his community. When physicians are seeing forty or fifty patients a day, he said, is just not sustainable. Dr. Stephens added that technology could be part of the burnout problem, explaining that the hours spent staring at a screen documenting patient encounters could also be a contributing factor as well. He said the ability to talk to a patient, do what you need to do for them, and move on, is getting lost somewhat due to technology. He said that AI could help with this where the doctor can talk to the patient, AI operates in the background, capturing his notes, diagnosis, etc., and when the physician walks out of the room all that other clerical work of entering information into a system is done.
Matthew followed up with a question on what is next regarding AI, Dr. Stephens said the low-hanging fruit in the near term is getting the real common images automated. The next step would have AI do the entire read with a physician being the second reader and signing off on it. For his specialty, Dr. Stephens sees a future where radiology operates more like a software company, where there is a feedback loop that updates the models as more imaging technology comes online. This could lead to one or several large companies and radiology becoming larger practices due to the cost of developing the models and maintaining them. Hopefully, all these advancements will improve patient care by getting them and their physicians more accurate results.
Matthew wrapped up and asked Dr. Stephens what he would say to young people considering medicine in general and radiology in particular. Dr. Stephens reassured us that technology was not going to replace radiology physicians. He followed up that if you are up to the challenge, the stress of going through medical school, and all it takes to become a physician, it will make you a better person. He encouraged those interested not to give up in their pursuit of a life working in medicine.
If you are interested in learning more about AI, follow the link to a recent episode of NOVA, the PBS science documentary series. They touch on some of the topics discussed by Dr. Stephens. AI Revolution on NOVA