Abstract No. 533 Deep Learning of Angiographic Signatures in Abdominopelvic Digital Subtraction Angiography

Abstract

Digital subtraction angiography (DSA) obtained at conventional intravascular locations assumes a typical appearance, or angiographic signature, unique to the location with similar imaging features across patients with conventional anatomy. Identification of these angiographic signatures is necessary for intraprocedural localization and planning and for appropriate post-procedural reporting and billing. The purpose of the study is to evaluate the use of deep-learning algorithms to identify angiographic signatures. An imaging database was developed based upon queries to an institutional PACS to identify abdominopelvic angiographic studies performed between 2010 and 2020, with correlate cross-sectional imaging performed within the preceding 150 days. Individual DSA series representing the angiographic signature of the abdominal aorta, celiac axis, superior mesenteric artery, inferior mesenteric artery (IMA), and right and left external iliac arteries were selected from the identified studies. In each DSA series, the sequence of images in which the parent vessel was opacified were labeled as diagnostic and used for the analysis. An Inception-ResNet v2 model was fine-tuned on a training subset for multiclass classification of angiographic signatures. Diagnostic performance was characterized by the precision, recall, and weighted F1 score across all angiographic signature classes and individually by class on a test subset. 647 unique angiographic sequences were included, comprising 3949 individual diagnostic images. The 5-fold cross validation multiclass classification performance was 0.917, 0.910, and 0.902 for precision, recall, and weighted F1 score, respectively. The angiographic signature classes were imbalanced with the percentage of the total images per class ranging from 29% for the celiac axis to 6% for the IMA. Performance by individual signature class tracked with the class imbalance. Recall performance, ranging from 0.992 for the celiac axis to 0.459 for the least represented IMA class, was most strongly influenced by class imbalance. Angiographic signature prediction is feasible and may be performed with high fidelity using deep learning algorithms.