Vision Transformers for AI-Driven Classification of Peripheral Artery Disease from Maximum Intensity Projections of Runoff CT Angiograms

• Published in: 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) pp. 3870 - 3872
• Main Authors: Salvi, Anish, Shah, Raj, Higgins, Luke, Menon, Prahlad G.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 06.12.2022
• Subjects: Angiography; Computed tomography; Deep learning; maximum intensity projection; medical image classification; peripheral artery disease; runoff computed tomography angiography projections; Sensitivity; Training; Transformers; Transforms; vision transformer
• DOI: 10.1109/BIBM55620.2022.9995337

Angled 2D views of computed tomography angiography (CTA) images (i.e. maximum intensity projections, MIPs) that have often undergone bone subtraction, allow radiologists to study vasculature unimpeded. Despite best efforts, subtle signs of peripheral arterial disease (PAD) can be underreported. Deep learning methods underpinned on regional attention can augment a radiologist's ability to identify PAD from these runoff CTA projections. A vision transformer (ViT) was built with 126 MIPs (from 2 Normal, 2 PAD patients) i.e. 66 Normal & 60 indicating PAD (aorto-iliac & distal aortic). The dataset was split into a training set of 46 Normal & 42 PAD and a validation set of 20 Normal & 18 PAD MIPs. The remaining MIPs (5 Normal & 9 PAD patients) were set aside for out-of-sample testing. Each MIP was resampled to 900x300 pixels before being split into three 300x300 segments along the subject's height. We implement a ViT that converts each MIP segment into 100 square patches, transforms each patch into numerical representations, and classifies each segment for PAD using these novel representations. PAD at the patient-level is determined by classification of segments across all MIPs identified as PAD. If the majority of segments across all MIP orientations of a given patient were scored as PAD, the patient was identified as having PAD. At the patient-level, we achieve 71% accuracy (78% sensitivity, 60% specificity) in out-of-sample testing. We report the first ViT in the literature capable of identifying PAD from runoff CTA projections.