Contrast-enhanced magnetic resonance imaging based calf muscle perfusion and machine learning in peripheral artery...

Peripheral artery disease (PAD) remains underdiagnosed and undertreated and is associated with an increased risk for adverse cardiovascular outcomes. Imaging provides an approach to identifying patients with PAD. However, the role of integrating imaging with machine learning to identify PAD patients and potentially assess disease severity remains understudied. A total of 56 participants, including 36 PAD patients with intermittent claudication and 20 matched controls, underwent contrast-enhanced magnetic resonance imaging (CE-MRI) calf muscle perfusion scanning. CE-MRI-derived dynamic muscle perfusion maps were developed to quantify alterations of the microvascular circulation in the calf muscles based on voxel contrast enhancement. These dynamic muscle perfusion maps categorized voxels as hyper-, iso-, or hypo-enhanced and were generated for the anterior (AM), lateral (LM), and deep posterior (DM) muscle groups, and the soleus (SM) and gastrocnemius muscles (GM). An unsupervised block-search algorithm was developed to identify heterogeneous regions of interest based on homogeneity. Machine learning methods were utilized to classify PAD patients from controls, with subgroup analyses performed based on lower extremity function and diabetes. The hypo-enhanced and hyper-enhanced voxel percentages obtained from the dynamic muscle perfusion maps were used to train a decision tree classifier to distinguish PAD patients from controls. The two-group classifier obtained a leave-one-out cross-validation (LOOCV) F1-score of 87.6 and 76.7% with hyper-enhanced and hypo-enhanced perfusion features averaged over all muscle groups, respectively. Hypo-enhanced perfusion features, a marker of microvascular perfusion abnormalities, classified PAD patients who completed a 6-minute treadmill walking test compared to those who did not, with an LOOCV F1-score of 67.6%. Using the same method, hypo-enhanced perfusion features differentiated PAD patients with diabetes versus those without with an LOOCV F1-score of 70.3%. In conclusion, CE-MRI derived measures of skeletal calf muscle perfusion can be used with a decision tree classifier to differentiate PAD patients from matched controls. Machine learning can also identify PAD patients with lower exercise capacity and those with concomitant diabetes. Machine learning and CE-MRI derived measures of the calf microcirculation could be of interest in the study of PAD and disease severity.