Segmenting Stenotic Regions from Coronary Angiograms using Mask Region-based Convolutional Neural Network

Coronary artery disease (CAD) is the primary cause of mortality and a key driver of healthcare expenses globally. Accurately segmenting stenotic regions from coronary angiograms is decisive in identifying and treating cardiovascular diseases. However, it is a challenging task for medical professionals to use X-ray Coronary Angiogram (XCA) due to the reduced signal quality, the existence of obstructive contextual elements, and various types of noise. Furthermore, handcrafted segmentation is arduous, laborious, and prone to inconsistencies and human errors. In this context, this research aim to develop an automatic stenosis segmentation system using a deep network. Initially, the input image is processed by Gaussian filters and the improved angiogram is filtered by Hessian-based Vessel Filtering (HVF) technique to increase the clarity of vascular components in the angiogram images. This study identifies the branch points (BP) in the angiograms based on the eigenvalues of the Hessian matrix. The proposed model employ a Mask Region-based Convolutional Neural Network (Mask R-CNN) to provide precise pixel-wise masks for every detected stenosis. The proposed Mask R-CNN includes (i) ResNet50 as the backbone network to extract significant attributes; (ii) Region Proposal Network (RPN) to identify possible Regions of Interest (RoIs) that may have stenosis; (iii) RoI Align to ensure precise alignment of the RoIs for improved mask prediction; and (iv) a mask branch to create a pixel-level segmentation mask for each RoI. The effectiveness of the model is assessed by applying an open-access ARCADE Phase 1 (Automatic Region-based CAD diagnostics using XCA images) dataset. The Mask R-CNN model achieves better results with 97.8% dice score, 92.9% sensitivity, and 96.6% specificity. Besides, it provides reduced standard deviation (SD) in the segmentation task with a 0.8% dice score, 1.0% sensitivity, and 1.0% specificity. These results shows that the Mask R-CNN model provides more reliable results for segmenting stenotic regions from angiography images.