LocNeXt: Fast Automatic ConvNeXt-based Coronary Ostia Landmarks Detection in 3D CCTA Images

This paper presents LocNeXt, a fast and efficient deep learning approach for automatic detection of coronary ostia landmarks in 3D Cardiac CT Angiography (CCTA) images. The method leverages the ConvNeXt architecture, incorporating modern design principles such as large kernels, inverted bottlenecks, and global response normalization. Rather than using vector field regression, our approach employs downsampled heatmaps with separate channels for each coronary ostium, transforming the detection problem into an efficient per-channel 3D classification task. The model was trained and evaluated on two public datasets (CAT08 and ASOCA) and achieved a median localization error of 7.04 mm for the quick version with smaller predicted heatmap size (LocNeXt-Q) and 5.24 mm for the full version, with inference times of 3.85 s and 24.29 s respectively. The method demonstrates robust performance despite the inherent anatomical variability of coronary ostia, particularly in the left coronary region, and proves to be data-efficient, requiring significantly fewer training images compared to previous approaches thanks to an intensive data augmentation pipeline.Clinical relevanceAccurate and automated detection of coronary ostia landmarks in CCTA images is crucial for various clinical applications, including pre-operative planning, diagnosis of coronary anomalies, and initialization of vessel segmentation and centerline tracking algorithms. LocNeXt provides a reliable, fast, and fully automated solution for this task, potentially improving workflow efficiency in clinical practice.