Optimal Deep Transfer Learning Models for Stenosis Detection in X-ray Angiography Images

Early detection of coronary stenosis is crucial in cardiology, allowing for timely diagnosis and medical intervention to prevent life-threatening complications like heart attacks and strokes. However, automating this diagnostic task is challenging due to background noise, non-coronary vascular structures, and multiple superposed branching points. In the past decade, Deep Learning (DL), in particular, Convolutional Neural Networks (CNNs), has shown remarkable success in various real-world applications, including medical imaging. Especially, Transfer Learning (TL), a technique where knowledge obtained from a large dataset is applied to different tasks with limited data, has become a standard in computer vision challenges with scarce training data. This chapter introduces a novel method for automatic coronary artery stenosis detection in X-ray Coronary Angiography (XCA) images. The method employs a pruning approach, where a pre-trained network is trimmed in an early convolutional block and connected to custom classifier layer. Additionally, the layers to be fine-tuned are optimally selected. Thus, this process reduces the number of transferred layers and trainable parameters. The proposed approach addresses a critical medical challenge. It demonstrates the potential of deep transfer learning for stenosis detection in XCA images, reaching an accuracy of 0.9612 with a Trimmed ResNet50 containing only 8.55 M parameters.