Adding AI to MRI improves return to play classification frameworks

Using an AI algorithm with MRI improves the performance of traditional return to play classification frameworks used with injured professional athletes, according to research presented December 3 at the RSNA meeting.

“MRI-based classification systems like BAMIC [British Athletics muscle injury classification] and MLG-R [Mechanism of injury, Location of injury, Grading of severity, and number of muscle Re-injuries] … especially when augmented by AI, enables improved prediction of time to return to play in professional athletes, facilitating evidence-based rehabilitation planning and return-to-sport decisions,” said presenter Sherif Elhennawy, PhD, of Saudi German Hospital in Cairo, Egypt.

Predicting when athletes can safely return to play after an injury is key, as allowing them back in the game prematurely makes them vulnerable to reinjury, while delaying return to play can translate into economic losses for the team, Elhennawy said. He explained that 30% of professional football injuries are muscle-related, and that 90% of lower limb muscle injuries involve muscles such as the gastrocnemius, rectus femoris, and long heads of the hamstrings. At least a third of professional athletes sustain at least one clinically significant muscle injury per season, he noted.

Elhennawy’s team investigated whether using an AI algorithm with MRI findings could improve the classification of sports-related thigh muscle injuries according to the British Athletics muscle injury classification (BAMIC) and MLG-R (Mechanism of injury, Location of injury, Grading of severity, and number of muscle Re-injuries). The study included 40 professional athletes who sustained thigh muscle injuries between 2018 and 2025, all of whom underwent 1.5-tesla MRI within 72 hours after injury.

The researchers defined the reference standard for time to return to play as the time from injury to full team training clearance (confirmed by the team physician and rehabilitation staff). They compared sensitivity, specificity, and area under the receiver operating curve (AUC) of BAMIC versus MLG-R, then assessed the performance of these systems with the addition of an AI model they built that consisted of supervised machine learning with MRI features.  

They found that, without the aid of the AI algorithm, a combination of the BAMIC and MLG-R classification systems was the most effective way to determine an appropriate return to play timeframe, especially regarding sensitivity.

But when the integrated classification model was enhanced with the AI algorithm, AUC improved from 0.984 to 0.993, sensitivity improved from 96.4% to 98.2%, and specificity improved from 94.4% to 96.7%.

“The enhanced sensitivity [imparted by the AI algorithm] reduces the risk of premature time to return to play,” he said.

He outlined the “performance hierarchy” of the time to return to play models — noting that AI-enhanced integrated showed the best overall performance and the integrated BAMIC plus MLG-R was the best traditional combination, while BAMIC alone was the best traditional single system and MLG-R alone offers “valuable anatomical detail,” — and concluded with a call for expanded validation studies with larger, multi-institutional cohorts and further refinement of deep-learning models for assessing appropriate time to return to play in professional athletes.