The Case for Broadening the Scope of AI in Radiology

By Mo Abdolell, CEO of Densitas

In the next couple of years, 50% of radiologists will use AI as part of their practice, the ACR Data Science Institute predicts. That makes sense given the explosion of AI applications hitting the market. As of October, the FDA listed nearly 700 AI devices that met its premarket requirements, with predictions that the number will only continue to rise. Most of those apps target radiology and within that category, the vast majority focus solely on diagnostics, which is a missed opportunity. AI can do so much more than that.

I already made the case that AI can drive healthcare quality, efficiency, equity, and sustainability but the question is, how? It requires more than putting our focus on AI for diagnosis and recognizing it can help with radiology operations in general, something I call “operational AI.” 

In the standard radiology protocol, a clinician requests an imaging procedure, assigns it to a radiologist or a technologist, the imaging data is acquired, processed, formatted, and then read and interpreted by the radiologist. That person then generates a radiological report, which guides the patient’s care. 

If operational AI is fully adopted, the radiologist can receive help with every part of that process, not only the diagnosis. For instance, AI can, and does, support workflow optimization by assigning a level of reading priority. It triages what a radiologist should read first, thereby increasing efficiency for the radiologist as an individual and the department as a whole. That has a trickle-down effect for everyone involved and ultimately benefits everyone, most especially patients. Other ways we can implement operational AI include the following:

• Using it to separate normal images from abnormal images
• Helping with quality assurance and evaluating whether an image has conformed to technical protocols, for instance in the case of a mammogram
• Grading and classification whereby AI provides image grading for structured reporting such as in the case of BI-RADS or PI-RADS 

All of what I just mentioned can be grouped under “pixel AI” because it’s image-based; however, there’s also text-based AI, which can speed up radiological report creation, a common cause for complaint among radiologists. With the help of AI, the entire process can become swifter and easier. 

AI in the form of natural language processing can convert unstructured text into a structured form to allow for automated information extraction or summarize findings into a pre-composed draft conclusion. In the case of computed-assisted reporting, AI can provide automated information in reports by linking to relevant clinical databases. Similarly, AI can support knowledge management by delivering contextual information from the EMR or dictated report and give guidance on how to categorize the patient and as well as offer management recommendations, for example in the case of dense breasts.  

The global market for medical AI applications is currently valued at $21 billion and is expected to reach $148 billion by 2029, according to MarketsandMarkets. Let’s do more than focus on diagnosis and employ AI in all of radiology’s operations to make it more efficient, effective, and ultimately improve patient care.