Artificial Intelligence in Radiology

Current AI can answer specific questions, but it is still a far cry from human-like intelligence. In most cases, questions evolve around predictions. In the diagnostic world of enterprise imaging, radiologists have been extensively trained to identify disease and are skilled at making efficient predictions—thus limiting the value of AI. In specific cases, algorithms could overcome the intrinsic limitations of human perception—such as skin cancer detection, diabetic retinopathy, or difficult-to-discern diseases on chest X-rays. Beyond these instances, the immediate future of AI lies more in aiding medically underserved areas.

However, AI is already far more valuable when its predictions are used to anticipate a need for resources. Examples include patient scheduling to reduce appointment no-shows; efficient allocation of IT and device resources through dynamic mapping; optimized assignment and triaging of radiologists’ cases; and automatic adjustment of smart display protocols to radiologists’ preferences in PACS systems. AI can guide and inform our decisions by generating new insights from tedious quantification tasks, such as Ejection Fraction computation in cardiac exams. Of course, AI is also adept at mining complex multi-dimensional data from multiple systems.

Medical Imaging-based AI

If an imaging product can help a referring physician determine the correct disease therapy — or significantly alter the treatment course through early detection of diseases such as lung cancer — it’s likely to have much higher value per study than if it simply helps a radiologist detect a lesion. Replacing a costly diagnostic procedure with a less expensive one will have a positive impact on cost reduction. (Example: replacing invasive, catheter-based devices with CT-based imaging phenotype of collateral ventilation.)

Imaging-based AI may also prove to be more accurate than genomics in predicting cancer for specific genetic subtypes, as we’ve seen in the fascinating brain imaging research in glioblastoma. Overall, identifying a patient’s risk of disease or preventing catastrophic events using imaging phenotypes will likely have a significant positive impact on a patient’s outcomes. Examples include the quantification of COPD to improve lung cancer risk assessment; the detection and screening of abdominal fat in relation to metabolic syndrome and cardiovascular disease; the prediction of aneurysm ruptures; and the detection of likely kidney damages in diabetic patients.

Perceiving Value in Artificial Intelligence

Payers, referrers, and radiologists all perceive value differently. In the near term, it’s likely that AI adoption will begin with applications embraced by all parties, for which AI will not be perceived as competing with our existing practices. AI tasks such as disease classification, which will require radiologists’ consent and directly impact their existing reading workflow, will take longer to be accepted. Segmentation tasks and mundane workflow-improvement tasks will initially be more appealing if they touch a large volume of cases and radiologists.

The consolidation of radiology groups will certainly help AI acceptance. In this initial phase, radiologists will need to be incentivized to contribute to the vendors’ AI ecosystems by providing quality annotated data and feedback. The interpretability of AI results from deep learning methods – a concern for physicians in general – will likely improve over time. Solutions such as local interpretable model-agnostic explanations (LIME) will evolve to analyze machine learning (ML) answers and point to the relevant source data.

Access to annotated data to train machine learning algorithms is essential for AI value creation. Data access will modestly improve in the U.S. as physicians’ bodies such as the ACR take aim at validating AI use cases. These groups are encouraging the provision of data that can be used for FDA submissions, as well as for competition between radiology groups for additional revenues. Longer term, any significant disruption to the regulatory process – coupled with the emergence of deep learning methods that consistently outperform humans – could dramatically increase the democratization of AI in radiology.

The infusion of AI in enterprise imaging is enabling ever more efficient workflows and data use. AI is expected to provide positive operational, financial, and productivity benefits before its full clinical value can be unlocked.

Learn more about enterprise imaging solutions from Change Healthcare.