Q&A: James Duncan on Big Data's Role in Medical Imaging

Artificial intelligence and big data have long held promise for revolutionizing the field of medical imaging with their potential for faster and more accurate readings. As the technology continues to advance, the field of biomedical imaging is closer than ever to realizing those aims. James Duncan, the Ebenezer K. Hunt Professor of Biomedical Engineering, Electrical Engineering & Radiology and Biomedical Imaging, recently edited a special issue of Proceedings of the IEEE on this topic. We spoke with him about the role of computers in medical imaging, how they can work with humans, and how this can lead to improved health care. 

How does big data fit in with medical imaging?

Humans bring their experience of seeing lots of these images and knowing the patterns and relationships and knowing overall medical knowledge, including demographics - how old is that person, what’s their body weight, what’s their medical history etc? The computer, especially those now using deep learning algorithms – must be told what to look at, such as images and perhaps specific demographic and other information. However, these algorithms can be trained with thousands - if not tens to hundreds of thousands - of these patterns and their very subtle differences. The human experts may be able to figure these out, but they need many, many looks at the views and the information and might need to go back and forth to look at previous data, while the computer can instantly recognize intricate patterns.

In both cases, they’re looking at the image data and trying to find, for instance, are there any lesions, where are they, how big are they? And then differentiate them - is it malignant or benign, or a dense set of tissue in and around certain parts of the breast? Machine or deep learning approaches can tirelessly perform these tasks. While human experts may still be better at diagnosing any one case, machines may ultimately be better at screening tasks.

How can human expertise and the power of computing be combined?

In many places, such as the UK, they require first and second readers [of the data], which can be tedious. Now, they’re talking about the computer being the second reader, so you still have that backup. Or, again, it could be that deep learning technology is better for screening, and when it finds difficult cases, that’s when the human gets involved.  

You edited a special issue of IEEE on all of this - tell us about that.

We have ten different groups of authors in different parts of the field trying to apply these techniques. Some are trying to recognize classes of tumors or of heart disease or something like that. Some are trying to improve the fidelity and quality of imaging, like ultrasound, positron emission tomography, or magnetic resonance imaging. The issue collects a set of these authors in these fields.

We tried to recruit the top people in these areas and we got some interesting contributions.  Maryellen L. Giger at the University of Chicago, who wrote one of the articles, has been one of the leaders in computer-aided diagnostics in breast cancer. And there’s Jeffrey Fessler at the University of Michigan, who has done a lot of work on image reconstruction and formation. 

It’s a nice collection of different things. One of the articles [by Li Shen of the University of Pennsylvania and Paul Thompson at University of Southern California] is on how you integrate imaging with genetics and genomics information. So there are some cutting-edge new things that are going on, and we’re all excited to see how it can all go forward and help the community and the patient population. 

It’s meant to look at the key issues in the use of deep learning/AI in medical imaging from a more technical perspective but still able to communicate to a broader audience. We wanted the people doing the research  to communicate not only to their own technical community, but how some of these topics affect others, including the doctors. 

How does this technology change things for patients?

The radiology community is very interested in this issue, but on the one hand are afraid that their efforts are going to be taken over by these machines. On the other hand, they want to be sure that they are involved with the development of these new approaches. From what I‘ve seen so far, the patients would likely benefit greatly no matter how the field develops, and will see it in terms of fewer false positives and fewer false negatives.

Also, in our radiology department here at Yale, there are several industrial collaborations and in this work, the algorithms have been able to find artifacts in images and distinguish them from the real diagnostic problems. So I think patients are going to see it in the long run as improved health care. In addition, in our field, one thing we think is going to happen is more personalized medicine - you can place yourself in a large spectrum and see more quantitatively where you are and physicians can develop targeted treatments and carefully quantify outcomes.

How long has it been used in medical imaging and where are we now with it? 

The field I’m in has been using machine learning approaches since the 1980s and 1990s and each year it has gotten more and more sophisticated. But the computational power and thinking weren’t there before, so it didn’t reach its potential. But now we’re at a new peak because we have the computer power (in part, due to the availability of graphical processing units) and the large amount of data that are becoming more and more available. However,  there remain many challenges that need to be overcome in the years ahead. Some of the key issues include limited amounts of image data related to specific medical conditions, as well as the need to fully interpret or clearly explain what enables some algorithms to be successful.