Detection of Arterial Occlusion on Magnetic Resonance Angiography of the Thigh using Deep Learning

Abstract

Magnetic resonance angiography (MRA) is an imaging tool used to evaluate arterial steno-occlusions in the lower limbs of patients with peripheral artery disease (PAD). This study aimed to train a deep learning method for the detection of arterial occlusions in the Superficial Femoral- and Popliteal Artery using radial maximum intensity projections (MIP) of contrast-enhanced MRA. A retrospective study was performed with 500 MRA exams included, using only the radial MIP of the thigh. Stenosis labeling was performed based on severity, considering only significant stenosis, and differentiating between no stenosis, focal stenosis, mid-length stenosis, and long stenosis. Class labels were combined to form four-class, three-class, and binary-class scenarios. An EfficientNet-B0 was trained and tested using a six-fold cross-validation for the right and left sides separately. The neural network (NN) achieved decent results with an area under the receiver operating characteristic curve (AUROC) of 0.917± 0.040 and accuracy of 0.851± 0.043 in the binary class case for the left side. The results degraded slightly for the three- and four-class cases and were overall minimally worse for the right side. The trained NN showed promising results in detecting arterial stenosis on MRA, which could potentially be a helpful tool for objectifying findings and reducing the workload of radiologists in the future.