In 2019 > 113 million people suffered from peripheral artery disease (PAD) whereas approximately 6.5 million people aged 40 and older in the United States have PAD [1,2]. PAD is associated with impaired blood flow in the lower extremities (LE), caused by atherosclerosis. PAD adversely impacts leg function and LE circulation, resulting in a significant risk for cardiovascular-related morbidity and mortality [3,4]. Timely risk assessment and early diagnosis can increase longevity and decrease the risk of cardiovascular events in PAD patients. Present risk identification methods could benefit from further improvements in order to reduce downstream effects caused by late-stage disease detection resulting in a higher rate of adverse outcomes [[5], [6], [7]].

Machine learning (ML) applications have been increasingly utilized in the medical domain and there is an utmost need to develop robust, dependable, and accurate ML models to improve risk stratification and clinical diagnostics [[8], [9], [10]]. Without exception, PAD diagnosis and pattern analysis from baseline to long-term patient characteristics has always been a subject of study [5,11,12]. ML algorithms applied with diverse data including electronic health record (EHR), imaging, genomics, clinical notes, and assessment scores from the patients' records have the potential to help clinicians rapidly risk stratify patients [7]. However, very few ML methods have been designed for the identification and risk assessment of PAD using magnetic resonance imaging (MRI) as a biomarker. The majority of PAD risk assessment and cardiovascular outcome analysis utilizing ML approaches were performed with electronic medical records (EMR) [[5], [6], [7],13], but not with imaging modalities.

PAD affects the microvascular circulation in the skeletal calf muscles resulting in impaired leg muscle function [[14], [15], [16], [17]]. Due to the irregular blood supply compared to healthy adults, PAD patients typically develop intermittent claudication pain. We hypothesized that prolonged ischemia in PAD patients manifests heterogeneous patterns of the calf muscle texture that can be quantified in non-invasive MRI scans to differentiate PAD patients from matched controls.

While various methods including local binary patterns [18], Gabor features [19], wavelets [20], and scale-invariant feature transform (SIFT) [21] exist for assessing image texture, Haralick's texture features are favored due to their simplicity, intuitive interpretations, and computational efficiency [[22], [23]]. In addition, textural analysis has high specificity, sensitivity, and reproducibility, making it a promising method for non-invasive imaging modalities including MRI [24]. Using Haralick's features from medical images has gained interest over time, and was utilized for the analysis of ultrasound and MRI images of liver lesions [25,26], classification of X-ray mammography [27,28], analysis of echocardiograms patterns as 2D image [29], coronary artery disease identification from iris images [30], COVID-19 prediction from CXR and CT images [31], identification of prostate cancer [32], cervical cancer [33] and brain cancer [34,35], all via MRI. Nevertheless, the use of MRI for identifying PAD has received less attention due to limited datasets for PAD groups.

In this paper, we evaluate the performance of Haralick's textural features that were derived from calf muscles contrast-enhanced MRI (CE-MRI) scans for the classification of PAD patients and matched controls by utilizing ML methods.