Seg2RefineNet: a novel DL-based framework for 2D CCTA image-based segmentation and 3D volume-based refinement

Computed Tomography Coronary Angiography is a non-invasive imaging technique widely used to assess structural abnormalities, blockages, or narrowing (stenosis) of coronary arteries, thereby aiding in the diagnosis and management of coronary heart disease. To assist clinicians in the assessment process, various AI-based methods have been proposed, both for 2D and 3D data, to accurately extract / segment the coronary arterial tree. This work aims to develop a novel two-stage hybrid segmentation method, Seg2RefineNet, to enhance coronary artery segmentation. The first stage employs a 2D spatio-frequency attention UNet, which results in the initial segmentation providing precise vessel boundary identification with high resolution. The second stage refines the segmentation using a 3D Attention-GAN, incorporating the inter-slice relationships within the 3D volume. As a proof-of-concept, this novel DL-based framework is evaluated on the largest publicly available dataset ImageCAS, outperforming the existing state-of-the-art methods by achieving a mean Dice score of 0.8313 and a Hausdorff distance of 12.95 mm. This hybrid approach effectively combines the strengths of both 2D and 3D models, setting a new benchmark for coronary artery segmentation.