The current state of AI in Radiology
By Steven L Blumer, Associate Medical Director of Radiology Informatics, UPMC
There has been much hype surrounding the potential of AI to disrupt healthcare and especially the field of radiology. Many radiologists have become concerned about being replaced by AI. Some medical students are fearful of pursuing a career in radiology for this reason. So how likely is this to happen and what is the current state of AI applications in radiology?
If we view the use of AI in radiology through the lens of the Gartner Hype Cycle, we have likely passed the peak of inflated expectations and started a descent into the trough of disillusionment. This is true because there is a misunderstanding about what AI can do in its current state within the field of radiology, especially when it comes to image interpretation. In this area, the current narrow AI algorithms can perform one specific task rather than fully interpret diagnostic imaging studies. For example, current AI algorithms can identify pneumonia and pneumothorax on chest x-rays as well as brain bleeds and pulmonary nodules on CT scans. Some of these narrow AI algorithms have performed as well or better than actual radiologists; however, this is not always the case.
Such algorithms are a long way from completely interpreting imaging studies. This ability requires incorporating a wide array of imaging and clinical findings and synthesizing this information to determine the most likely diagnosis or diagnoses. These “human-like” abilities are what many people think about when they believe that AI has the potential to replace radiologists. However, these “human-like” capabilities require a higher level of AI known as Artificial General Intelligence (AGI). AGI or strong or full AI applications in radiology are still nascent and will likely take many years to fully see any impact on the radiologist workforce.
The current state of the technology is such that AI is a useful adjunct to radiologists in image interpretation and allows radiologists to improve their efficiency and quality of care in other parts of the imaging lifecycle.
The use of narrow AI in other areas of the imaging life cycle is currently further along and has proven more beneficial to radiologists. These applications can make radiologists more efficient and improve patient care. This has become especially important as the volumes of imaging studies continue to increase dramatically.
Image acquisition is an important part of the imaging lifecycle. AI algorithms currently in use can perform QA on imaging studies and alert the technologist when it may be necessary to repeat a CT scan or sequence from an MR examination based on poor image quality. Some algorithms can take exams of poor diagnostic quality and use AI to generate exams of diagnostic quality. Such algorithms allow for shorter scan times for MR examinations and lower radiation dose exposures to patients during CT scans. Similarly, there are also AI algorithms that have been developed for contrast enhanced studies which can augment contrast enhanced images. This allows radiologists to administer lower doses of contrast agents to patients on CT and MR exams.
As sub-specialization within radiology practices increases and demand for rapid report turnaround times for large volumes of radiology studies also goes up, it has become paramount for the right study to be routed to the right radiologist at the right time. AI-enabled workflow orchestration systems are now in use that route radiology studies to radiologists based on their areas of expertise within radiology as well as the urgency of the study. This helps ensure that each study receives the highest level of subspecialty expertise and is read as quickly and efficiently as possible. In addition, such systems can help ensure a more even distribution of imaging studies to each radiologist and can help alleviate physician burnout. Furthermore, there are also AI algorithms that can detect potential urgent findings such as brain bleeds and prioritize these studies at the top of the reading worklist to ensure that a radiologist views these studies as quickly as possible, potentially decreasing the time to life-saving treatment.
NLP is a branch of AI that allows unstructured data to be transformed into structured data. This technology has been especially useful in its applications to radiology reports. For example, when a pulmonary nodule or thyroid nodule is mentioned in a report, current AI algorithms recognize these findings and generate current recommendations for their clinical management that can be automatically included in radiology reports for clinicians to reference. There are also algorithms that use NLP to search reports for imaging findings such as pulmonary nodules that require follow-up and flag them in the EMR so that providers can see these studies and ensure adequate follow-up. NLP algorithms can also search reports and find laterality mismatches and sex/gender mismatches so that these can be corrected, preventing potential adverse patient outcomes such as wrong-site surgeries. NLP algorithms have also been developed to automatically generate the impression section for radiology reports based on imaging findings mentioned earlier in the report. This ensures that important findings do not get overlooked in the impression section of the report and can help the radiologists save time in generating report impressions and allow them to read more studies in a shorter amount of time.
It is also important that radiology departments are staffed adequately to handle imaging volumes. Current AI algorithms can predict volumes of imaging studies for each institution on a given day so that the correct number of radiologists and staff are assigned to work during those shifts. Such algorithms can predict how many studies each radiologist can read based on the average number of studies that they have read in the past. This allows for better pairings of radiologists not only to anticipate volumes but also with other colleagues to ensure equity among the group. For example, these applications may help ensure that a slower reader is not paired with another slower reader and is instead paired with a faster reader so that the work is done more efficiently.
As demonstrated above, most of the narrow AI algorithms currently available do not have the ability to replace radiologists. The current state of the technology is such that AI is a useful adjunct to radiologists in image interpretation and allows radiologists to improve their efficiency and quality of care in other parts of the imaging lifecycle. However, we have not reached the point of the development of AGI algorithms that will potentially replace radiologists which may never happen. But as Dr. Curtis Langlotz of Stanford University has said, “Radiologists who use AI will replace radiologists who don’t.”