A pair of respected researchers - one from Harvard Medical School, and the second from Stanford University and the University of California, San Francisco - have just published a very thoughtful and quite readable review article in the New England Journal of Medicine (NEJM) describing the current and future state of affairs in the utilization of artificial intelligence (AI) algorithms and their application to clinical practice. Currently, radiologists use AI tools not only to aid in the detection of abnormalities and other conditions on medical images, but also in workflow triage and in their quantification of findings from the images. So, what is AI's future?

The use of AI in radiology, they report, has shown great promise in detecting and classifying abnormalities on plain radiographs, computed tomographic (CT) scans, and magnetic resonance imaging (MRI) scans, leading to more accurate diagnoses and improved treatment decisions. Curiously, they make little mention of AI in connection with the detection of abnormalities on mammograms.

Currently, although penetration of AI in the U.S. market is estimated to be only 2%, they report, the readiness of radiologists and the potential of the technology indicate those numbers will grow substantially.

The authors believe that even though the Food and Drug Administration (FDA) has approved more than 200 commercial radiology AI products, substantial obstacles must be overcome before there will be widespread clinical use of many of these approved AI products. For example, performance of many radiologic AI models can worsen when they are applied to patients who differ in character from those used for model development. This is known as "data set shift", i.e., the shift from data used to train a machine-learning model to data encountered in the real world. They advocate moving from evaluations of AI's effectiveness centered on the stand-alone performance of models to evaluations centered on the outcomes when these algorithms are used as assistive tools in real-world clinical workflows. Such will offer a better understanding of the effectiveness and limitations of AI in clinical practice and establish safeguards for effective clinician–AI collaboration.

AI in radiology, they argue, will require development of new "foundation models" which are defined as those AI models that serve as a starting point for developing more specific AI models. Foundation models are trained on large amounts of data which can be fine-tuned for specific applications, such as detecting lesions or segmenting anatomical structures.

As for the future, they describe what some writers of science fiction may have imagined a decade or so ago: application of AI algorithms to evaluate both the medical imaging as well as relevant clinical information, resulting in the production of a complete radiologic report for the radiologist, a patient-friendly report with easy-to-understand descriptions in the preferred language for the patient, recommendations regarding a surgical approach that are based on best practices for the surgeon, as well as evidence-based follow-up suggestions and tests for the primary care provider. A world in which radiological findings become even more influential than in today's clinical practice. Wow!

Those interested in the current and future of AI in medical imaging should read this excellent NEJM review article to gain a better perspective on the progress, the challenges, and the opportunities in the development of radiologic AI and their clinical application.

The narrow focus of existing AI solutions on interpretation of individual images in isolation has contributed to the limited penetration of radiologic AI applications in practice.