Vision Transformer in stenosis detection of coronary arteries

Abstract

About 380 thousand people in the United States die each year because of coronary artery disease. About 2 in 10 deaths from CAD happen in adults less than 65 years old. Early detection of blocked arteries can prevent serious surgeries and decrease the number of deaths. Coronary angiography, as one of the methods for diagnosing the state of the arteries, is a complex and invasive procedure. Neural networks can be used to support physicians in the diagnostic process and improve stenoses detection effectiveness. In this paper, we compare different variants of the Inception Network and the Vision Transformer in a stenosis detection task that, in our case, is a binary classification problem. The data set used in the experiments consists of small fragments extracted from coronary angiography videos. Furthermore, we analyze the impact of different percentages of arteries in fragments without stenosis on model performance and show that the data set configuration plays an important role. As the fragment images are small, we analyzed the Sharpness-Aware Minimization together with the Visual Transformers. We employ explainable AI methods to understand the differences in classification performance between selected models. Our results show that convolutional neural networks generally perform better than transformer-based architectures, but perform slightly worse if the Visual Transformer model is supported by the Sharpness-Aware Minimization method. While increasing the percentage of arteries in non-stenosis class, the F1 median decreases by 2.3, 5.4, and 2.2 for the Inception, ViT, and SAM-ViT models. The F1 mean decreases by 3.4, 4.8, and 2.3. The SAM-ViT is more stable compared to other models and based on the F1 mean outperforms other models. This indicates the usefulness of ViT in medical applications, especially in the analysis of coronary angiography.