Editorial
Br J Sports Med: first published as 10.1136/bjsports-2018-099999 on 22 November 2018. Downloaded from http://bjsm.bmj.com/ on 9 January 2019 by guest. Protected by copyright.
Artificial intelligence in sports medicine Data scientists at the University of British
Columbia and at Simon Fraser University
radiology: what’s coming?
in British Columbia are working on many
applications that automate the measure-
ments of structures in medical images,
William Parker, Bruce B Forster such as liver lesions, spinal vertebrae and
bone lesions.
With all of the potential benefits of
Artificial intelligence (AI) is the new kid and there are no algorithms that have been AI, care will need to be taken to ensure
on the block. Every doctor and her/his approved without being strictly supervised patient safety. Radiologists will need to
medical student is talking about it. It seems by a trained physician. Radiologist over- become trained and proficient in how
like the answer to all of our problems. sight is essential to keeping patients safe. these algorithms come to their conclu-
How will we overcome antibiotic resis- AI is also starting to branch into sions in order to best interpret their
tance? AI will figure it out. What is on the supporting point of care ultrasound. results. The algorithms will need to have
X-ray? AI will know. Why is your partner AI optimised time gain compensation a certain level of transparency as well, in
mad at you? AI should be great for that. control, AI simplified needle tip enhance- order to show the user how the result was
But what is AI exactly, and what will be its ment and AI automated heart rate analysis, calculated, what the sensitivity and speci-
impact in sports medicine? and automated segmentation and differen- ficity of the algorithms are and when the
AI is the theory and development tial diagnosis list for lesions are starting to system identifies images that it struggles
of computer systems which are able to become available for portable ultrasound to interpret. Some of the demos shown
perform tasks that normally require systems. The authors feel as though these by start-up companies have partially
human intelligence, such as visual percep- features will make ultrasound easier for transparent colours overlaying the images
tion, speech recognition, decision-making clinicians to use at the bedside. Having that becomes more intense in the region
and translation between languages.1 In said this, an understanding by the clini- of the abnormality, called a heat map.
order for a machine to become intelligent, cians with respect to how to interpret Some systems are even able to learn
it needs to learn. Enter machine learning. ultrasound and its AI components will be from mistakes from radiology input, to
Machine learning is a branch of AI based essential.7 improve for future studies.
on the idea that systems can learn from It is our opinion, however, that AI will The future of Radiology will involve
data, identify patterns and make decisions change the specialty of radiology for the a significant amount of AI. But as these
with minimal human intervention.2 better, and as an extension, medicine in systems develop, we must remember who
Machine Learning is not a new concept general. But only certain algorithms will we are serving first and foremost: the
in computer science circles but was popu- have a positive impact, and there will be patient. As sports medicine physicians,
larised by Dr Geoffrey Hinton, a computer many others that miss the mark and will having a basic understanding of what AI
scientist from Toronto Ontario, who pass in the wind. is and how it is being used by your local
brought the technology into the limelight Up to 42% of mistakes made by radiol- radiologist, is important. The point should
with his work on ImageNet, a computer ogists are due to detection or perceptual ALWAYS be to improve patient care, and
program (called a neural network) that was errors, which may be where AI could help.8 patient outcomes, with the ultimate goal
able to identify the contents of 1.2 million This means that a finding was present in of improving quality of life. Doing all of
images, with only a few errors.3 4 the images, but the radiologist failed to see this while taking advantage of exciting
Naturally, Radiology is a medical it. Interestingly however, once a finding is new technology is definitely a bonus.
specialty in which machine learning has seen by a radiologist, only 3% of all errors
the potential to be highly disruptive; it subsequently are due to a lack of knowl- Funding The authors have not declared a specific
could assist radiologists with increasing edge of what the finding is or what to grant for this research from any funding agency in the
efficiency, decreasing errors and improving do next.9 Thus, if AI could help decrease public, commercial or not-for-profit sectors.
the health and well-being of patients.5 detection errors in Radiology, the radiol- Competing interests BBF has an equity position in
Start-up companies have already created ogist could be better prepared to help a private imaging clinic in Vancouver BC. WP conducts
algorithms that calculate bone density for patients. Many companies are focusing on research and has an equity stake in a company focused
prediction of osteoporosis on CT, identify on machine learning for Radiology.
this very thing.
basic anatomic structures on knee MRI A recent orthopaedic study found that Patient consent Not required.
and detect fractures on CT scans.6 These even though ‘inter-observer agreement Provenance and peer review Not commissioned;
companies have graphic websites and between radiologists [was] higher than externally peer reviewed.
impressive demos, which create consid- among orthopaedists in the evaluation of © Author(s) (or their employer(s)) 2018. No commercial
erable excitement. However, radiologists chondral knee lesions by MRI’, radiolo- re-use. See rights and permissions. Published by BMJ.
and clinicians also need to be careful. gists have a significant amount of interob-
Though many companies are advertising server variability (kappa of up to 0.78).9
their work in AI, not many have been For example, AI could help to standardise To cite Parker W, Forster BB. Br J Sports Med Epub
approved by health regulators (less than the way radiologists measure certain ahead of print: [please include Day Month Year].
a dozen have been approved by the FDA) chondral lesions, measure the displace- doi:10.1136/bjsports-2018-099999
ment of fractures, tumour response to Accepted 8 November 2018
Radiology, University of British Columbia, Vancouver, treatment, grade muscle injury and assist Br J Sports Med 2018;0:1–2.
British Columbia, Canada with many other measurements. This doi:10.1136/bjsports-2018-099999
Correspondence to Dr William Parker, Radiology,
could improve the interobserver vari-
University of British Columbia, Vancouver, BC V6T 1Z4, ability between radiologists and stan- References
Canada; w illiam@alumni.ubc.ca dardise the information for clinicians. 1 www.google.com/d efine artificial intelligence
Parker W, Forster BB. Br J Sports Med Month 2018 Vol 0 No 0 1
Editorial
Br J Sports Med: first published as 10.1136/bjsports-2018-099999 on 22 November 2018. Downloaded from http://bjsm.bmj.com/ on 9 January 2019 by guest. Protected by copyright.
2 SAS. “Machine Learning: What It Is and Why It be measured. https://m edium.com/@BalintBotz/ 8 Bruno MA, Walker EA, Abujudeh HH, et al.
Matters.”. www.sas.com/en_c a/insights/analytics/ a-few-thoughts-about-chexnet-and-the-way-human- Understanding and confronting our mistakes:
machine-learning.html#machine-learning-importance. performance-should-and-s hould-not-be-measured- the epidemiology of error in radiology and
3 Hinton GE, et al. ImageNet classification with deep 68031dca7bf. strategies for error reduction. Radiographics
convolutional neural networks. Neural Information 6 "Zebra Medical Vision.". Medical imaging & ai. www. 2015;35:1668–76.
Processing Systems Foundation, Inc 2010;60:84–90. zebra-med.com/. 9 Cavalli F, Izadi A, Ferreira AP, et al. Interobserver
4 The Royal Society Fellow. Geoffrey Hinton. https:// 7 Dickie K. Clarius - Portable Ultrasound Scanner. reliability among radiologists and orthopaedists in
royalsociety.o rg/people/geoffrey-hinton-1 1624/. “Current approaches to artificial intelligence in medical evaluation of chondral lesions of the knee by MRI. Adv
5 Botz B. A Few Thoughts About CheXNet - and the imaging.”. clarius.com/e ducation/white-papers/artificial- Orthop 2011;2011:1–4.
way human performance should (and should not) intelligence-medical-imaging.
2 Parker W, Forster BB. Br J Sports Med Month 2018 Vol 0 No 0
