Deep Learning-Based Stenosis Segmentation in X-ray Angiography: Vision Transformers vs. CNNs

Abstract

This thesis compares ten deep learning models for automated stenosis segmentation in X-ray coronary angiography (XCA) images to support coronary artery disease (CAD) diagnosis. The study evaluates both Transformer-based and CNN-based architectures using the ARCADE dataset under a unified training framework. Results show that the Transformer model SegFormer MiT-B3 achieved the best performance, producing more accurate and anatomically consistent vessel segmentation than CNN baselines. The findings highlight the advantage of Transformer architectures in capturing long-range vessel relationships, while lightweight models such as MiT-B0 and MobileNetV2 showed strong efficiency for real-time clinical use. Future work should include stenosis severity measurement, spatiotemporal modeling, and validation on multiple datasets.