Coronary angiography remains the gold standard imaging modality for the examination and treatment guidance of coronary artery diseases during percutaneous coronary intervention (PCI). However, the two-dimensional projection of three-dimensional details in coronary angiography only allows for a rough estimation of the degree of vessel blockage and its location [1]. Intravascular imaging, specifically intravascular optical coherence tomography (IVOCT), has emerged as a complementary modality for guiding PCI [2]. It offers intravascular assessment to aid in pre-PCI clinical decision-making (e.g., stenosis localization and stent selection), immediate evaluation of PCI outcomes (e.g., stent deployment quality) and long-term PCI assessment (e.g., stent failure/in-stent restenosis) [3]. Lumen segmentation in IVOCT images is a crucial step for obtaining several essential intravascular measurements, including the severity, location and length of vessel stenosis, as well as the degree of stent malapposition and restenosis burden. Manual lumen segmentation is infeasible due to the large number of images within a single IVOCT pullback, which typically consists of 300-400 images. Consequently, automating the segmentation process becomes imperative to ensure prompt and accurate assessment.

Previous studies have employed various image processing techniques for the automated lumen segmentation of IVOCT images. For instance, Nam et al. [4] and Menguy et al. [5] utilized depth intensity profiles of polar IVOCT images to determine lumen coordinates. Tsantis et al. [6] employed a Markov random field model to extract lumen borders, while their extended work [7] utilized fuzzy c-means clustering to identify the lumen region. Several publications proposed active contour models based on different approaches, such as level sets [8], snakes [9] and geodesic active contour [10]. Additionally, Xu et al. [11], Modanloujouybari et al. [12] and Zhang et al. [13] performed graph-based segmentation, each differing in design, including aspects like image features (intensity gradients [11], RGB color [12] and gradient vector interaction [13]), cost function (Dijkstra's algorithm [11] and piecewise model with Gaussian kernel function [12]) or incorporating a star shape prior [13]. Some research groups [14], [15], [16], [17] primarily applied mathematical morphological operations to define the lumen border. These operations encompassed binarization, opening or closing, often combined with additional operations. Examples included catheter/guide-wire elimination using area and orientation angle [14], anisotropic linear-elastic mesh surface fitting [15], Sobel filter [16] and blood artifact removal through the uniqueness of vascular connected regions [17]. Furthermore, Fuzzy systems [18], random walks solver [19], and L-mode interpolation [20] have also been adopted for IVOCT lumen segmentation.

In recent years, deep learning has demonstrated its potent capability to push the boundaries of medical image segmentation [21]. Yong et al. [22] constructed a linear-regression convolutional neural network (CNN) that processed cropped polar IVOCT images to generate corresponding radial distances of lumen. Tang et al. [23] introduced N-Net, which was built upon the U-Net architecture but integrated multiscale input layers to fuse original IVOCT images of various resolutions into the skip connections. Gharaibeh et al. [24] employed a pretrained SegNet (with random field processing) to refine segmentation outcomes. Huang et al. [25] devised the Residual Squeezed Multi-Scale Network (RSM-Network) based on the U-Net framework, enhancing it with pyramid feature extraction (PFE) structures and residual attention (Residual-SE) structures. Similarly, Balaji et al. [26] modified the U-Net model by incorporating capsules and dynamic routing.

Traditional image processing techniques yield reasonable results but often necessitate specific preprocessing to address common artifacts like blood swirls and shadowing from the catheter, guide wire and stent struts. In contrast, modern deep learning, driven by data, inherently manages such artifacts while achieving satisfactory or even superior outcomes. Most of the cited deep learning studies [23], [24], [25], [26] applied pixel-wise classification to distinguish the lumen region from the background in IVOCT images. This technique is widespread in general image segmentation, where the objects of interest can assume arbitrary shapes. However, these deep learning pixel classification techniques often require post-processing to eliminate outlier pixels. Conversely, Yong et al. [22] acknowledged the circular or convex nature of the lumen and employed CNN regression to accomplish lumen segmentation. This regression-based segmentation did not necessitate further post-processing to remove outlier pixels. Their approach involved inputting narrow cropped images and generating individual radial distance points to construct the lumen incrementally, one step at a time, consequently limiting the speed necessary for real-time PCI. We hypothesized that inferring all radial points simultaneously from the entire image could offer superior speed and lumen continuity. Thus, in this paper, we developed a one-shot image regression network architecture, designed to infer sets of polynomial coefficients to represent the lumen border in a single pass. We anticipated that this approach will encourge the network to learn the lumen border in connected segments rather than individual radial points, thus enhancing boundary smoothness and avoiding discontinuities. The proposed network was implemented using a large dataset and verified on diverse clinical scenarios.