Automated Artery Detection and Stenosis Classification in CTA Using Deep Learning for Peripheral Arterial Disease

Peripheral Arterial Disease (PAD) is one of the most important cardiovascular diseases worldwide for several million humans, associated with substantial morbidity, mortality, and healthcare resource utilization. Computed Tomography Angiography (CTA) is a high-quality imaging modality in PAD diagnostics, which directly highlights the vasculature having high resolution. However, manual interpretation of large-scale CTA data is hindered by human errors and is time-consuming. In this study, we introduce an automated deep learning model for detecting the main arteries and stenosis severity along with occlusions inference in lower limb CTA scans. We use Faster R-CNN with a ResNet-101 backbone driven by a custom loss function to achieve good artery localization and reduce false positives. The evaluation of our model is done by experts on a 67,850 CTA slices dataset from eighty patients. Experts manually annotated a subset of 13,439 slices for use in training and evaluating the model. Post-processed with HU-based vessel analysis and intersection-over-union (IoU) filtering, stenosis classification was more refined. In this way, the proposed method resulted in an overall accuracy of 98.79%, with an AUC of 0.9534 for stenosis classification. Deep learning integration with automated stenosis quantification will thus present a tool for robust AI-driven diagnosis of PAD that will bring about a clinically viable solution towards enhancing diagnostic accuracy, reducing workload, and improving patient outcomes.