Machine learning to improve sports injury detection and performance
Thursday, 31 August, 2023
Niamh Belton is a PhD student in the UCD School of Medicine and a member of (opens in a new window)ml-labs, a Science Foundation Ireland centre for research training in machine learning. She is currently working on automating the diagnosis of knee injuries from MRI data using deep learning.
Last year Ireland rugby captain Johnny Sexton missed the game against Argentina because of his. Leinster flanker Dan Leavy had major surgery - and was ruled out for most of the season - because of his. Connacht’s Stephen Fitzgerald had to retire on medical grounds because of his. Knee injuries truly are the bane of the professional rugby player’s life. Indeed, of any sports player’s life.
“So the quicker the diagnosis and the quicker you can intervene with appropriate rehabilitation, the more likely it will be to have better patient outcomes,” says Niamh Belton, whose work looks at how machine learning - a branch of artificial intelligence - might improve the speed and accuracy of knee injury detection.
Currently, the “gold standard” is the MRI scan.
“The problem with this is once the MRI is taken you then need an expert, a radiologist, to review it and give the diagnosis. But radiologists can be very busy and this can take time.” Diagnoses can also be “subject to a level of human error”.
What Niamh wants to do is “develop an automated tool that can look at the MRI and give injury diagnosis within seconds”.
Such Computer Aided Detection (CAD) systems are built using machine learning techniques and deep learning models, or systems.
“Machine learning is essentially teaching the computer to learn by itself,” explains Niamh, before giving the example of how it might learn to detect the dreaded anterior cruciate ligament (ACL) tear.
“You would input the data - MRIs of the knee - into the machine learning algorithm and it automatically learns features and finds patterns in the data and it learns how to detect ACL tears by itself. Then in practice, you can just feed this model a new MRI, and it will output the probability of an ACL tear, with no human intervention whatsoever.”
She is not suggesting that these algorithms will replace radiologists.
“The whole point of the computer aided detection system is that it would work as an aid to the radiologists. So the MRI would be input into the system and it would output the probability of having detected an injury. And it might also output a sort of heat map highlighting the areas of the MRI where the model found some sort of abnormality. That might allow the radiologist to make a more informed decision and speed up diagnosis time.”
Along with acting as a support for the radiologists, “the accuracy can be good enough that it can work by itself. It is possible to have an MRI machine in a stadium while a football match is ongoing. It could output an injury probability there and then”.
The accuracy of injury detection models depends on the quality of the data inputted in the first place.
“In real life situations in hospitals, you might have to deal with poor quality data so the model needs to be robust to poor quality data.”
When Niamh looked at MRNet, the open source knee MRI dataset released by Stanford University, she found MRIs containing incorrect or no anatomical information.
“Basically you want to remove them from the data set because they will just harm the model’s training process when it is trying to figure out how to learn to detect ACL tears.”
Anomaly detection - detecting anomalies in datasets - has become “a big field” in computer science. Niamh found her own way around the problem.
“Using a very small subset of normal MRIs we managed to train the model to automatically detect the bad data.”
Her “overall aim” is to develop an automated way to generate radiology reports for musculoskeletal injury diagnosis. The first step is injury detection - and she will be working with data from professional rugby players and marathon runners, looking not only at knee injuries but at hamstring injuries and more.
“It would be great to conduct a user study of our injury detection model with radiologists to ensure that the tool is helpful to them. Would they be happy for this system to be implemented into a hospital or clinic? I would like to go more in that direction because you want your research to have an impact.”
The next part of her research - funded under the Irish Research Council’s Ulysses scheme - sounds like it would be equally of interest to coaches.
“We are moving into muscle architecture analysis because we hope using deep learning to perform muscle architecture calculations might give us some information on risk of re-injury or return to play. And once we do that, we could also include that information on the radiology report.”
Niamh will travel to Marseille University shortly to work on this project.
“We are collaborating with a group over there that has access to MRI data and DTI (Diffusion Tensor Imaging) data and we are going to try to automate some muscle architecture analysis.”
She particularly enjoys the interdisciplinary nature of her research.
“It’s great to be across two schools and to develop a more thorough understanding of machine learning techniques, but also to be able to apply them in the medical fields.”
Listen to the (opens in a new window)podcast