Non-Interpretive AI Offers High-Impact Solutions

While much-touted image-interpretation algorithms might one day provide autonomous care, it will be many years before they deliver the results we see with non-interpretive AI today.

When we think about AI in radiology, most radiologists think of machines interpreting images — and look forward to a day when AI will support physicians by detecting subtle, critical findings while freeing up time to become more patient-facing. Although this image-interpretation use of AI is an exciting part of our future, it is not our most pressing need now and might cause us to overlook something equally exciting: non-interpretive AI.

Much of the current development of image-interpretation AI is focused on creating algorithms to identify a specific finding or make a specific diagnosis, as you can see from the list of the FDA-cleared algorithms on the ACR Data Science Institute™ (DSI) website. But for the most part, AI is not critically needed for image analysis since radiologists now accomplish the tasks that AI tools can solve. Realistically, these image-interpretation AI tasks will not be rapidly implemented in our practices. It will take time.

Non-Interpretive AI Solves Current Problems

In the near-term, analysis in Academic Radiology and coverage in other medical imaging news sources explain why we may be better off looking first to non-interpretive AI. Non-interpretive AI holds the promise of helping radiology practices solve current, large-scale problems, including:
• Predicting patient wait time
• Predicting patient no-shows
• Improving image quality
• Automating worklist prioritization

While non-interpretive algorithms that tackle efficiency and productivity issues may not be as sexy as AI rapidly interpreting high volumes of scans, they have greater potential to be rapidly developed and adopted. Rapid deployment and implementation is feasible because non-interpretive algorithms (for the most part) do not suffer the same issues as image-interpretation algorithms — incorporation into physician workflow, FDA-clearance, and providing clear value. 

In general, non-interpretive algorithms can be readily incorporated into existing workflow. Since these algorithms do not interpret images, they are less likely to face costly hurdles before receiving FDA-clearance and can enter the market more quickly. Non-interpretive algorithms have the potential to address each element of the imaging value stream from the moment an imaging study is considered by the ordering physician through final billing and subsequent imaging follow-up — delivering quick results.

Use Cases for Non-Interpretive AI

Recognizing the value non-interpretive AI can bring to radiology, the ACR DSI Non-Interpretive Panel has been working to identify and build use cases to define these non-interpretive problems that are highly solvable by AI. The panel comprises six sub-panels, each taking a different radiology practice user’s perspective: ordering providers, patients, technologists, radiologists (both in the reading room and in the procedure room), business administration, and the general population.

To date, our sub-panels have been successful in identifying ways for AI to improve workflow management through non-interpretive AI. The subpanels have published 17 AI use cases for radiology in five categories.

Business Facing
Automated follow-up program
Computerized auto-coding of reports with real-time dictation feedback
Intelligent routing of exams for optimal performance
Predicting patient no-shows for radiology appointments
Predicting volume to optimize staffing
Reconciling discrepancies on insurance payments

Patient Facing
Chatbots to answer radiology-based procedure patient questions (breast imaging)
Computer-aided translation of radiology reports (thyroid ultrasound) to layperson language
Information delivery on incidental findings (pulmonary nodules)
Produce multi-media reports that are easier to understand
Update patients on wait times

Population Health Facing
Decreasing variability in follow-up recommendations (incidental thyroid nodules)

Reading Room Facing
Automated cross-sectional co-registration
Prioritization of exams on the worklist
Radiology and pathology report correlation
Virtual transcriptionist/dictation assistant

Technologist Facing
Detecting image quality in medical imaging

Concluding Thoughts

At the end of the day, AI applications for both imaging analysis and efficient workflow management have the potential to benefit radiology. The two uses reinforce one another. But since non-interpretive AI brings value to radiology departments as soon as it is put into place, it has the added benefit of providing an impact right away. While image-interpretation algorithms will one day be common, it will be many years before they offer us the kinds of results we can achieve now with non-interpretive AI tools.

 

Alexander J. Towbin, MD | chair of Radiology Informatics, and associate chief of Radiology, Clinical Operations and Radiology Informatics at Cincinnati Children's Hospital Medical Center, Cincinnati, OH