_{Published March 1, 2022}

Oganes Ashikyan

Associate Professor
Department of Radiology, Musculoskeletal Imaging Section
UT Southwestern Medical Center

Hillary W. Garner

Assistant Professor of Radiology
Mayo Clinic Florida

Some of you may remember the time when your voice recognition stubbornly transcribed “pulmonary edema,” even though you clearly said “bone marrow edema.” One of the co-moderators of our 2022 ARRS Annual Meeting Sunday Session, “Practical Applications of Computational Science in Musculoskeletal Imaging,” was a second-year resident when voice recognition was first implemented in his department. His faculty and co-residents were divided on whether to welcome or resist this new technology. The issue was not about young versus old, or tech savvy versus not, but about whether or not the technology was ready for prime time. Initially, voice recognition was not as accurate as advertised. It took time for the software to mature to its current, more robust form.

Years later, the moderator was taking an online introductory artificial intelligence (AI) course, during which the technology behind voice recognition was explained. Contrary to his early assumption, the software was not trying to transcribe phonemes into the exact words. Instead, it was listening to chunks of speech within sentences and assigning probabilities to what was being said, ultimately displaying the highest probability word(s) related to the overall context. Large amounts of voice and context data were required for the software to be able to achieve high accuracy and allow for appropriate probability models for various users in different subspecialty settings.

Today, we hear about AI and other computational technologies achieving unbelievable feats. We are approached by salespeople who already have FDA-approved software that can perform tasks that were only achievable by the human mind just a few years ago. Some of these computer and data science solutions will be able to stand the test of time. Other solutions will seem incredibly promising but inevitably fail. Regardless of which solutions persevere, very few people in health care have a firm understanding of how these technologies work and what limitations they may have. As physician leaders, we need to narrow this knowledge gap to help contain costs and to better serve and protect our patients. To achieve these goals, we need to improve our ability to judge available solutions, including their potential benefits and drawbacks, as well as their probability of wide adoption and success. Having a basic understanding of how things work in computational science is therefore becoming as necessary as being able to read, write, and do basic arithmetic in today’s rapidly advancing high-tech world.   

During “Practical Applications of Computational Science in Musculoskeletal Imaging,” you will hear from fellow physicians on various topics that integrate computer science and musculoskeletal imaging, including automated evaluation of arthritis, use of AI in speeding up MRI acquisitions, automating measurements of bone loss, and performing fat versus muscle mass measurements. Furthermore, we hope to answer some popular questions in the realm of musculoskeletal AI, such as where we are today regarding a completely automated solution in fracture detection on radiographs and what it means for a software solution to be FDA approved, as opposed to FDA cleared. You will also hear about potential pitfalls and biases that AI may introduce. We hope that you will join us in learning more about this exciting and continually evolving aspect of musculoskeletal radiology.