Conference coverage from RSNA 2014
A comprehensive collection of articles, videos and more, hand-picked by our editors
Will your next radiologist be a supercomputer? There are a growing number of companies evaluating complex image processing algorithms based on pattern recognition. Will one of them invent a computer that is capable of diagnosing medical conditions by reviewing X-rays, CT scans, MRIs, ultrasounds and other imaging studies? I believe the potential for automated diagnostics is there. This level of clinical decision support is already being incorporated into certain imaging studies such as mammography, but it may become the standard of care in the not-too-distant future.
We may find that our next radiologist is a computer.
In medical school, physicians are taught algorithms and clinical treatment pathways. The science of medicine is combined with the art of medicine, based on experience and intuition. Computers are capable of adaptive learning, so it is theoretically possible to "teach" a computer to become better at interpreting imaging studies and make fewer errors over time. We are seeing major advances in computing and artificial intelligence technology that incorporate natural language processing with adaptive learning and big data analytics. We saw how the IBM Watson supercomputer beat a human contestant on Jeopardy. Perhaps we will witness a similar type of standoff between a top radiologist and a supercomputer.
I predict future radiologists will have a level of clinical decision support that will significantly reduce diagnostic errors when they read films. For example, when reviewing a CT scan of the abdomen, the computer will identify patterns of suspected abnormality and alert the reviewer that there is a 95% probability of a ruptured appendix (based on specified findings), a 75% probability of an abscess (based on specified findings) and a 99% probability of diverticular disease (based on specified findings).
The radiologist might be so focused on the appendix that he misses the fact that the CT scan also detected a suspicious lesion on the right adrenal gland, so the computer could alert the radiologist to this finding. The computer may also alert the radiologist to degenerative bone changes on the spine. These latter findings are some common examples of diagnostic errors and misses that occur in medicine because people tend to be laser-focused on the acute problem -- the appendix -- and may miss other findings that are less obvious.
In the United States, we are about to see thousands of new patients enter the healthcare system. Most hospitals and health systems are not prepared for this influx of new patient volume, so there is an immediate need for technology-driven solutions that will enable radiologists to provide higher-quality, higher-efficiency care at lower costs.
Even if highly capable automated diagnostics tools were available today, they would not be incorporated into routine practice right away. Before such algorithmic, adaptive learning technologies are incorporated into routine clinical care, clinical researchers will conduct studies to evaluate how supercomputing clinical decision support solutions reduce medical errors and improve diagnostic accuracy.
After all, these solutions will be expensive and they will require high-powered computing equipment and intensive IT support to install and maintain them alongside PACS and EHR solutions. In some countries, a cost-effective analysis will be necessary to justify reimbursement by a national single payer. As data emerge on the clinical and economic benefit of such solutions, we may find that our next radiologist is a computer that makes a preliminary diagnosis, followed by a radiologist who confirms the diagnosis and signs off on the case.
Joseph Kim is a physician technologist who has a passion for leveraging health IT to improve public health. Dr. Kim is the founder of NonClinicalJobs.com and is an active social media specialist. Let us know what you think about the story; email email@example.com or contact @SearchHealthIT on Twitter.