A public cardiac CT dataset featuring the left atrial appendage

TL;DR

This work tackles the scarcity of public, high-resolution segmentations for the left atrial appendage (LAA) and neighboring structures by creating an open-source CCTA dataset built on the ImageCAS collection of 1000 scans. The authors combine TotalSegmentator-derived whole-heart labels with a Neural Unsigned Distance Field (NUDF) based LAA segmentation and a distance-field–based refinement approach for the pulmonary veins (PVs), culminating in anatomically coherent 10-class labelmaps. Key contributions include high-quality LAA surfaces that capture fine anatomical features, refined CA and PV labels, and a scan-quality list to aid downstream analyses, enabling robust LAA morphology analysis and CFD studies. The dataset supports broader stroke etiology research and provides a benchmark for comparing LAA morphology descriptors across methods, despite limitations such as unknown cardiac phase and limited metadata.

Abstract

Despite the success of advanced segmentation frameworks such as TotalSegmentator (TS), accurate segmentations of the left atrial appendage (LAA), coronary arteries (CAs), and pulmonary veins (PVs) remain a significant challenge in medical imaging. In this work, we present the first open-source, anatomically coherent dataset of curated, high-resolution segmentations for these structures, supplemented with whole-heart labels produced by TS on the publicly available ImageCAS dataset consisting of 1000 cardiac computed tomography angiography (CCTA) scans. One purpose of the data set is to foster novel approaches to the analysis of LAA morphology. LAA segmentations on ImageCAS were generated using a state-of-the-art segmentation framework developed specifically for high resolution LAA segmentation. We trained the network on a large private dataset with manual annotations provided by medical readers guided by a trained cardiologist and transferred the model to ImageCAS data. CA labels were improved from the original ImageCAS annotations, while PV segmentations were refined from TS outputs. In addition, we provide a list of scans from ImageCAS that contains common data flaws such as step artefacts, LAAs extending beyond the scanner's field of view, and other types of data defects.

Figures (5)

• Figure 1: Left atrial appendage morphologies. Examples of a chicken wing, cactus, cauliflower and windsock morphology from ImageCAS following the classification system of DIBIASE2012531.
• Figure 2: Example of the provided data. Surfaces represent the fused labels, while orthogonal cross-sections of the LAA and a volume rendering are used for visual quality control. Note that in this example, the LAA (green) and the PVs (dark-red) nearly overlap.
• Figure 3: Comparison of non-cropped, randomly selected scans from ImageCAS. Top) Our segmentations. Middle) TSTOTAL segmentations. Bottom) combined segmentations. Shown are files 992, 957, 949, 572, 443, and 239 from ImageCAS. The voxelized, blocky nature of the TotalSegmentator LAA segmentations results from the low image resolution (1.5 mm spacing) on which they were generated, combined with the use of marching cubes for rendering.
• Figure 4: Expert annotation vs. NUDF segmentation. Left) Lowest NUDF Dice score (81.5%) on the 10 manually annotated cases from ImageCAS. Right) Highest NUDF Dice score (92.1%). It is apparent that locating the ostium is challenging and contributes the most to loss of Dice score.
• Figure 5: Comparison of TSTOTAL segmentations and postprocessing of pulmonary veins. Left) LA with the original TotalSegmentator segmentation of the PVs. The blocky surface stems from the low resolution image the segmentation is created on. Right) LA with the refined PVs. Note the intensity change of the contrast enhanced blood pool due to step artefacts caused by the scanner acquisition.