Fully Automatic Severity Grading of Lower Limb Arterial Blockage based on Plantar Perfusion Imaging

Lower limb arterial blockage (LLAB) has recently been measured by camera based on plantar perfusion, offering a new method for the early prevention of peripheral arterial diseases (PAD). In this paper, we conducted a simulation experiment that applied five pressure levels to the left calf to simulate different degrees of LLAB and collected plantar video data from 20 healthy participants. We benchmarked four machine learning methods for LLAB severity grading based on plantar perfusion images generated by remote photoplethysmography. In addition, we propose the ResNet18-CL method, which combines ResNet18 with contrastive learning, to address issues related to the difficulty in distinguishing severities (i.e. different pressure levels). The experiments showed that ResNet18 outperforms conventional machine learning methods in LLAB grading. Compared to ResNet18, ResNet18-CL improves accuracy and recall by 5.98%, precision by 6.58%, and F1 score by 7.36%. This approach provides a timely, costeffective, and efficient solution for the severity grading of LLAB.