Radiologists do not just look at images, the same for artificial intelligence (AI) in radiology. The second AIMed specialty event – AIMed Radiology concluded at Ritz-Carlton, Chicago yesterday (19 June). The one-and-a-half-day conference was held back to back with AIMed Cardiology and consisted of a series of hands-on practical workshops, as well as panel discussions chaired by distinguished professionals and leaders that were live-streamed across the World. 

In one of the sessions, Tessa Cook, Assistant Professor of Radiology, University of Pennsylvania; Nina Kottler, Vice-President of Clinical Operations at Radiology Partners; Ross Filice, Associate Professor and the Chief of Imaging Informatics in the Department of Radiology at MedStar Georgetown University and the Chief of Imaging Informatics for Medstar Medical Group, and Kevin Seals, Interventional Radiology Fellow at University of California, San Francisco Medical Center, had injected audience with information on the impacts of AI and machine learning (ML) beyond image interpretation, detection, and decision making. 

Cook began by illustrating the benefits of having a system whereby she does not have to manually retrieve and sift out information from patients’ electronic medical records (EMRs) to look at prior examinations and results to determine what need to be done next. She said the opportunity which AI had rendered include the capability of integrating existing and new information and able to deploy them to track patients’ progresses over time. 

Tessa Cook presenting on the impact of AI and ML on radiology

Overcome variabilities to provide better follow-up recommendations 

Kottler synchronized with Cook and added that since clinicians and radiologists no longer meet in the reading rooms to discuss recommendations, radiology reports become more crucial than ever. However, there are many variabilities in terms of language, definition, and what should be done as follow-up. If these differences could be minimized in some ways, it will add tremendous value in the provision of care. 

A way to do so is to introduce a PDCA cycle or rather what known as the Plan (i.e., identify your problems), Do (i.e., test potential solutions), Check (i.e., study results) and Act (i.e., implement the best solution) guidelines, to ensure there is a standard which all clinicians in the same institutions can follow. However, the approach was not scalable, and clinicians have to repeat PDCA cycle many times for different recommendations, so Kottler and her team had developed recoMD, an ML-generated decision support system. 

Nina Kottler sharing the recoMD platform

The recoMD interface is able to sit on top of other software that the clinicians are using to generate their reports. It mainly provides clinicians with follow-up recommendations for a particular patient. If the clinician does not agree with the given recommendations, he or she can write down what they disagree about, to help improve this ML tool. At the same time, to avoid AI Blackbox, clinicians will also be told of the assumptions made by the system that lead to the given recommendations. Variabilities were found to have lowered just weeks after rolling out the recoMD. 

Improving radiology reports 

Filice’s presentation also touched upon radiology reports. He highlighted the importance of pathology in radiology reports and the attempt to have the pathology report automatically come back to radiologists before they decide on what to do with the patients. Filice’s team made use of an existing ML model, a generalized one that can understand language with the underlying architecture being recurrent neural network so that it can comprehend the meaning of words next to each other. He then fed it with 200,000 radiology and pathology reports to understand the prose of the realm. 

Ross Filice speaking on stage

The model is now able to perform binary classification problem, informing radiologists whether the radiology report matches with the pathology report or not. It also has a high specificity of more than 90% and an intentionally induced sensitivity rate of 60-70% to avoid false positive. Filice believes the most efficient part of ML is it can be rapidly trained on its own based on feedback and be customized based on the needs of individual radiologist to improve performance. 

A smart speaker that tells you about your tools

For Seals, he is tapping on smart speakers which assist intervention radiologists to better perform their procedures. He said, often when clinicians are in their sterilized clothes, they become a little restricted in what they can do. Their interaction with technology is compromised because they are not able to hold onto their phones or to type anything on their computers. On the other hand, the exponential growth of technology also means that there are many new devices and tools being found in the procedure rooms. 

Kevin Seals sharing his project on stage

Sometimes, these new devices may come with a user manual and the technologists in the room may assist by following through the guidelines given by the manuals. Other times, the companies that manufactured these tools will send a representative to help with the radiologists. Either way, conflicts may arise due to interests and achieveing a common understanding. Because devices and tools come in diverse forms, shapes and sizes, radiologists may not be able to know exactly which tool is ideal for that particular blood vessel diameter. 

Seals’ team generated a large database of tools that intervention radiologists commonly use. From sheaths, catheters, coils, to plugs, embolic and stents and passed it onto an ML algorithm before the output was fed into a smart speaker. Now, all radiologists have to do is to ask the smart speaker for the answer. Seal believes the system could be scaled up, and eventually be linked to an inventory, so on top of a recommended tool, radiologists will also be armed with the information of what kinds of tools are made available. 

You may wish to re-visit the AIMed Radiology seminar and panel discussions here. Do follow us on Twitter, Instagram, Facebook and Youtube for more event updates.

Author Bio
synthetic gene empathy chinese artificial intelligence data medicine healthcare ai

Hazel Tang

A science writer with data background and an interest in the current affair, culture, and arts; a no-med from an (almost) all-med family. Follow on Twitter.