Deep learning-based classification of superficial femoral arterial lesions: a pilot study

This study evaluated the feasibility of detecting peripheral arterial lesions in plane-reconstructed lower extremity CT angiograms using object detection algorithms.

We retrospectively collected 1,241 contrast-enhanced lower extremity CT images from patients with peripheral arterial disease. One-stage (YOLOv5: v5s, v5m, v5l, v5x) and 2-stage (Faster R-CNN) detectors were used to classify stent, stenosis, and occlusion. A one-by-one comparison between manual test image annotations and algorithmic detection results was conducted to evaluate model performance and errors. Performance was evaluated by mean average precision (mAP@.5) and precision-recall curves.

Among YOLOv5 models, v5l showed the highest overall accuracy (77% mAP@.5). While stent classification was excellent (≥ 96.9% mAP@.5 in YOLOv5 and 99.8% in Faster R-CNN), classification accuracies for stenosis (53.8%-58.7% in YOLOv5 vs. 37.2% in Faster R-CNN) and occlusion (69%-80.9% in YOLOv5 vs. 67.7% in Faster R-CNN) were moderate. Stenosis was frequently missed, resulting in high false-negative rates. Occlusions at arterial bifurcations were often not detected, and stent edges were misclassified as occlusions. Overfitting emerged in some YOLOv5 models beyond 75 epochs.

This pilot study supports the feasibility of applying object detection algorithms as a preliminary step toward developing clinical decision support tools for peripheral arterial disease. Further refinements, including additional training data and more granular lesion annotation, are essential for improved classification of stenosis and occlusion.