Using deep learning to predict muscle function based on anatomical MRI scans from the OAI and SOMMA datasets

Motivation: Estimating physical performance including muscle function is conventionally done by having a subject complete physical tasks. MRI-based estimates leveraging deep learning could complement such measures. Goal(s): To investigate the feasibility of predicting measures physical performance including muscle strength from MRI scans of the leg using deep learning. Approach: We used large MRI databases (OAI and SOMMA) to train a neural network for classification into high or low physical performance. We also tried the method on a small prospective cohort. Results: We obtained over 70% accuracy for estimating high or low muscle function, indicating high predictive power. Impact: We demonstrate the feasibility of predicting muscle function from anatomical MRI scans using deep learning, leveraging advances in deep learning and musculoskeletal MRI databases that include functional measures. Such MRI-based predictions could complement conventional methods for estimating muscle function.