Unleashing the Promise of Artificial Intelligence in Radiology

“Benefits of using AI can include one or more of making the task faster, easier, less costly, more accurate, or more consistent/repeatable. These benefits can also be across individual users or large clinical groups; AI brings uniformity to derived physiologic and anatomic measurements across a potentially large number of human performers for a given use case. Without it, individuals are left all doing their own thing. AI makes it easier for a group of people to collectively do the right thing easily.”
--- IHE. AI Interoperability in Imaging. 2021.[1]

It’s been five years since artificial intelligence (AI) pioneer Geoffrey Hinton stunned the radiology profession by saying: “We should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists.”[2]

Actually, it’s completely obvious that hasn’t happened. In fact, there are now more practicing radiologists in the US and growing concern over shortages.[3] What AI is changing is how radiologists work.  “AI is quite important in radiology,” said Felipe Campos Kitamura, MD, Head of Innovation in Diagnostic Operations at Dasa, Brazil's largest integrated healthcare network that currently takes care of the journey of more than 20 million people a year. “More and more, people are convinced that radiology practices will change because of it.”

Indeed, the field is slowly recognizing the value AI brings to the profession with its ability to find patterns that humans are not able to see easily in image data. With this capability, AI can help detect lesions, take on tedious tasks like segmenting structures and help triage case criticality.[4] It can help automate routine tasks such as chest X-ray rotation, saving thousands of interface clicks; prioritize workflow; and even reduce no-shows by mining years of appointment data. [5]

 And yet, said David Mendelson, MD, a radiologist at Mt. Sinai in New York, “I think we’re at the very beginning of AI, even though it’s a few years old right now.” Dr. Kitamura, considered one of the leading experts in imaging AI in the world, agreed, saying we are still in the “infancy” of AI.

They’re right. A 2020 survey of 1,427 radiologists in the US found that just a third reported using any type of AI, although 20 percent of practices planned to buy AI tools in the next one to five years. The survey found that most were using AI to help enhance interpretation, particularly of intracranial hemorrhage, pulmonary emboli, and mammographic abnormalities. Barriers to its use include concerns over inconsistent performance, potential decrease in productivity, and lack of reimbursement, the authors found.[6]

Nonetheless, Dr. Kitamura is optimistic. “I think the real value will be in the next few years when people start to deploy AI in a more robust way and it’s completely integrated into the workflow,” he said. And that will depend in part on the culture of the practice.

Overcoming the Barriers

If AI is to be successful, it must be integrated into the medical imaging ecosystem.
--- American College of Radiology. AI Interoperability in Imaging White Paper. 

One of the biggest barriers to integrated AI lies in the sheer number of vendors and the complexity of many of the products they provide, said Dr. Kitamura. “Evaluating all those offerings would take months or longer,” he said. “Then they have to be integrated into the existing infrastructure, requiring not only IT support but significant training.”

But before organizations can embark on their AI journey, a proper imaging informatics infrastructure need to be in place first. Think of it as needing the cake before you can put the cherry on top. The underlying infrastructure, such as the PACS, must work well before AI is added.

“What’s needed is a single vendor with a ‘plug-and-play’ solution that works across platforms,” he said, “one where you don’t need to go to a different workstation in a different physical location; you don’t need to open a different window. It’s everything in the same place.” Any institution moving into AI today, he said, should consider a vendor-neutral tool.

Data integration and analytics have been cited in industry surveys and expert interviews as one of the major trends reshaping imaging services and influencing the future of healthcare. Vendor Neutral Archives (VNAs) have an important role to play in connecting existing systems to share data between clinical applications, as well as connecting clinical teams themselves, while helping enable data to generate actionable insights and artificial intelligence (AI) driven decision support. But while AI-based operational insights algorithms are retrospective and hence a natural to run off a data archive, AI-based clinical imaging applications often processing exam data before a radiologist reads the exam. Thus, it is essential to have a way for these AI-Based imaging applications to integrate with the radiology PACS and integrate with the reading workflow. 

“We recognized quite some time ago that the only way practices would truly embrace AI is if it worked seamlessly with the imaging devices or PACS software they already had without requiring a completely new system,” said Peter Eggleston, GE Healthcare’s Global Product Marketing Director for radiology IT.  “That’s why we developed Open AI Orchestrator, designed to seamlessly integrated clinical applications into the radiology reading workflow of our GE Healthcare PACS customers, using the PACS interface radiologists are already familiar with.”

It’s also important to individualize how AI is used for the specific practice and hospital, Dr. Kitamura said. “Each place may have different protocols and different processes, and this will influence whether your solution works or not.” For instance, Dasa, which employs more than 2,000 radiologists, uses a rule-based system to customize PACS to assign images depending on the time of day and subspecialty. With AI, Dr. Kitamura said, more criteria can be added to the decision making to help narrow the differential diagnosis.

Another issue is bias. “Bias in AI means it works for one group of people but not for another group,” he explained. This could be any minority in the data set. For instance, without children in the dataset it won’t work for a pediatric population.[7] Same with gender, ethnicity, and race. “All AI models have bias,” he said.

To assist developers in understanding and minimizing data bias, GE Healthcare’s Edison Developer Program gives developers access to a rich set of AI capabilities for data traceability, curation, annotation, model training and inferencing. This set of services allows data to be traced during the development of an algorithm. These AI capabilities are just one set of the more than 100 Edison services available to GE Healthcare and ultimately third-party developers.

Another barrier to AI has been the capacity to generalize, Dr. Kitamura said. “Just because the model has been trained to work with images from a small training set, doesn’t mean it will work as well in a generalized patient population,” he said.   

The issue of generalization, however, may be improved with the use images acquired in a diverse set of medical imaging equipment. Dr. Kitamura and his colleagues at the Mass General Brigham Center for Clinical Data Science (MGH-CCDS) developed of an AI algorithm to assess the FLAIR sequence of the brain MRI to determine abnormalities.[8] He and his team tested their approach on ~10,000 brain MRI exams from two institutions on two different continents, demonstrating that the diversity of MRI scanner models can aid in generalizability across the institutions with respect to MR scanner models. [9]

As they noted in their paper: “Given the complexity and variability of brain MRI protocols, we are confident that similar techniques could be applied to other forms of radiological imaging.” His practice recently deployed it in a controlled environment and expects to see improved workflow, he said.

Realizing the Power of AI in Radiology

The current focus with AI is on helping with diagnostic capacity and its ability to support radiologists in high profile, acute conditions such as intracranial hemorrhage, pulmonary embolism, and pneumothorax, said Dr. Mendelson. Being able to assess these conditions in minutes makes a big difference in the patient’s outcome, he said. “We all know that if we have two physicians reviewing a piece of information, you’re less likely to have errors,” Mendelson said. “Now AI can provide a second set of eyes.”

The ability to triage [10] for the radiologist for cases that are more likely to be urgent is particularly helpful, he said. “Particularly in high-volume practices, (this helps) the most acute illnesses rise to the top for interpretation.”

Artificial intelligence is about much more than aiding in diagnosis, however. “AI is beginning to prove quite valuable in optimizing workflow,” said Dr. Mendelson.  That includes patient direct scheduling and scheduling templates based on the amount of time an exam will take as well as determining how many patients should be seen in an hour for that modality.

It can even help address one of the biggest problems facing the field: burn out. One study found that 54 percent of radiologists reported feeling being burned out or depressed.[11] Not only have workloads steadily increased over the last 20 years due to the increased utilization of imaging, the work is getting more complicated.[12],[13]

The American College of Radiology recommends improving radiologist efficiency as one way to reduce stress and burnout, and there are studies demonstrating that AI is an important tool in helping meet that goal.[14],[15] “We need radiologists to work smarter, not harder,” said Dr. Kitamura. “And that’s where AI can help.” For instance, one study found a nearly 30-minute decrease in total interpretation time across all readers of MRI scans when using an analytics-driven worklist.[16] In a 100-radiologist practice, that translates into 50 saved hours a day.[17]

“We can use AI in the entire imaging cycle,” said Dr. Kitamura, “from the scheduling to the ordering to the communication of the results.” Also tracking productivity and errors. “Some of the most important applications may be things you wouldn’t even notice,” he said, “such as reconstructions and post processing.”

 One thing radiology practices will notice is how AI can reduce no-shows, a significant problem in radiology, with rates of 6.5 percent or higher.[18] Unused slots mean lost income and disrupted workflows. But an AI-powered scheduling software like GE Healthcare’s Smart Scheduling uses up to 40 different data factors to understand patient probabilities for missing appointments, potentially reducing no-shows by up to 70 percent.[19]

Another important function, said Dr. Mendelson, is the ability of AI algorithms to speed the processing at the modality.  “In many cases you can shorten the acquisition of the image data yet get equal image quality, if not enhanced quality, using AI algorithms on top of the standard data collection,” he said.

Just as helpful would be the ability of AI to aid in recognizing the “normal.” For example, take stroke. “In terms of workflow, there’s actually a greater value in recognizing the normal cases even without a specific diagnosis because you can get patients in and out of the ED much faster” stated Dr. Kitamura. This is particularly important as the volume of ED visits in the US continue to increase.[20] Plus, in the ED every case is a STAT case. With stroke, for instance, time is brain.

As for that concern that AI might one day replace human radiologists “We are far from that outcome,” said Dr. Mendelson. “I do think most radiologists now are recognizing that as a supplement to their diagnostic acumen AI can be very valuable in certain situations.” Maybe it’s time, then, that we think of AI not as artificial intelligence; but as augmented intelligence – augmenting the human intelligence of the doctor.

To learn how GE Healthcare is incorporating AI into the PACS workflow, visit our website.

 



[1] IHE International. AI Interoperability in Imaging. https://www.ihe.net/uploadedFiles/Documents/Radiology/IHE_RAD_White_Paper_AI_Interoperability_in_Imaging_Rev1-0_PC_2021-03-12.pdf
[2] Geoff Hinton: On Radiology. November 24, 2016. Available at: https://www.youtube.com/watch?v=2HMPRXstSvQ
[3] Smith G, Funk J. Quartz. AI has a long way to go before doctors can trust it with your life. June 4, 2021. https://qz.com/2016153/ai-promised-to-revolutionize-radiology-but-so-far-its-failing/
[4] Loria K. Putting the AI in Radiology. Radiology Today. 19(1):10. Available at: https://www.radiologytoday.net/archive/rt0118p10.shtml
[5] Loria K. Intelligent Upgrades. Radiology Today. 22(5):18. Available at: https://www.radiologytoday.net/archive/rtJJ21p18.shtml.
[6] Allen B, Agarwal S, Coombs L, et al. 2020 ACR Data Science Institute Artificial Intelligence Survey. JACR. 2021. DOI:https://doi.org/10.1016/j.jacr.2021.04.002
[7] Patient population for any AI application is based on what the AI app is regulatory cleared or approved and intended for use.
[8] This work is not a product developed in collaboration with GE and is also not distributed by GE Healthcare (confirm that this statement is true).
[9] A Deep Learning–based Model for Detecting Abnormalities on Brain MR Images for Triaging: Preliminary Results from a Multisite Experience. Romane Gauriau, Bernardo C. Bizzo, Felipe C. Kitamura, Osvaldo Landi Junior, Suely F. Ferraciolli, Fabiola B. C. Macruz, Tiago A. Sanchez, Marcio R. T. Garcia, Leonardo M. Vedolin, Romeu C. Domingues, Emerson L. Gasparetto, and Katherine P. Andriole. Radiology: Artificial Intelligence 2021 3:4.
[10] The capability to triage depends on the regulatory cleared/approved intended use of an AI application.
[11] Kate L. Medscape National Physician Burnout, Depression & Suicide Report 2019. January 16, 2019
[12] Harolds JA, Parikh JR, Bluth EI, et al. Burnout of radiologists: frequency, risk factors, and remedies: a report of the ACR commission on human resources. J Am Coll Radiology 2016; 13:411–416.
[13] Parikh JR, Wolfman D, Bender CE, Arleo E. Radiologist Burnout According to Surveyed Radiology Practice Leaders. JACR. 2020;17(1):78-81
[14] Harolds JA, Parikh JR, Bluth EI, et al. Burnout of Radiologists: Frequency, Risk Factors, and Remedies: A Report of the ACR Commission on Human Resources. JACR. 2016;13(4):411-416.
[15] The AI Effect. MIT Technology Review. Available at: https://www.technologyreview.com/hub/ai-effect/.
[16] Wong TT, Kazam JK, Rasiej MJ. Effect of Analytics-Driven Worklists on Musculoskeletal MRI Interpretation Times in an Academic Setting. AJR Am J Roentgenol. 2019 Feb 26:1-5.
[17] Saving is based on a specific case study but can vary depending on different hospital practice
[18] Harvey HB, et al. Predicting No-Shows in Radiology Using Regression Modeling of Data Available in Electronic Medical Record. J Am Col Radiology. 2017;4(13):1303-1309.
[19] Results based on single pilot site, results not typical and cannot be guaranteed. GE Healthcare makes no guarantees concerning the impact of the use of Smart Scheduling on customers' facilities.
[20] National Center for Health Statistics. Health, United States, 2019: Table 37. Hyattsville, MD. 2021. Available from: https://www.cdc.gov/nchs/hus/contents2019.htm.Available at https://www.cdc.gov/nchs/data/hus/2019/037-508.pdf