Introduction

With an aging population, musculoskeletal disorders associated with aging are expected to increase in prevalence. This includes sarcopenia, the loss of muscle mass, strength, and performance. Muscle strength and performance are often measured by having the subject complete tasks such as walking or pressing a device. However, such reliance on patient cooperation can introduce noise in the measurements. Here, we developed machine learning algorithms to predict physical function from MRI. Neural networks were trained to make such predictions based on large databases containing MR images and various physical function measures. After training the network on acquired images, we tested the network on scan data acquired at our own institution.

Methods

As training data, we used images and physical function measurements from the Osteoarthritis Initiative (OAI) and the Study of Muscle, Mobility, and Aging (SOMMA)^1,2. We obtained 15 bilateral axial MRI slices in the mid-thigh from each of 3,515 OAI participants and 835 SOMMA participants. Each cohort also collected measures of leg strength (isometric strength chair for OAI, Keiser leg press for SOMMA), walking speed over short and long distances (20 m and 400 m for OAI, 4 m and 400 m for SOMMA), and the chair stand test. For our network architecture, we used a ResNet18 neural network, tuning it to our data set by training the weights of the last network layer. OAI and SOMMA data were used separately to train two separate networks with this same methodology, splitting each dataset into 90% for training and 10% for validation, with the validation accuracy used as the performance metric. For the predicted outcome, we trained the network to predict whether the subject had a high or low physical function for each functional measure, defined as being above or below the median value of the cohort. For further investigation, the OAI network was also trained separately on males and females from the database, still using the median of the combined cohort as the classification threshold. Upon the completion of the training, we used the best performing network to predict the physical function of 8 healthy volunteers scanned at our own institution. For these subjects, self-reported exercise (“yes”/”no”) was used as the physical performance measure to be predicted and the imaging was designed to replicate the OAI imaging settings for the thigh. Sample inputs are shown in Figure 1, with image parameters in Table 1. The network was implemented in PyTorch. Training was performed over 50 epochs using an Nvidia GeForce GTX 1080 GPU, with training taking about 7 hours for the OAI dataset and about 2 hours for the SOMMA dataset. Other training parameters included using the SGD optimizer with a learning rate of 0.0001 and a momentum of 0.9, using the Cross Entropy Loss function, and a batch size of 8. The network structure is shown in Figure 2.

Results

Of the four physical performance methods, the leg strength measure was found to yield the best predictive accuracy for both the OAI and SOMMA networks, with over 70% validation accuracy for both datasets (OAI results shown in Figure 3). For the whole cohort as well as for the male- and female-specific networks, the predictive success rate exceeded the baseline of the relative size of the larger class within each cohort (e.g., for the female cohort, 60% were in the “low strength” class, a baseline exceeded by the network). Less predictive accuracy was obtained for other functional measures such as walking speed and chair stands. Saliency maps did not show an obvious pattern of heavily weighted regions, but may indicate an emphasis on the tissue periphery (Figure 4). For the prospective scans at our own institution, the OAI network successfully classified 5 out of the 8 subjects.

Discussion

The results demonstrate that a neural network based on the ResNet18 architecture can classify subjects into categories of high or low leg strength based on thigh MRIs with an accuracy of over 70%. While not conclusive, the observed pattern in the saliency maps of weighting the tissue periphery may point to the network using thigh size as a determinant. This is further supported by the reduced accuracy above baseline when stratifying by gender. Further investigation to control for body size, for instance by stratifying by BMI, is ongoing.

Conclusion

Axial thigh images can be used to classify subjects into having high or low physical function with over 70% accuracy using deep neural networks. This approach could complement standardized methodologies for assessing function.