Thoracic Aortic Three-Dimensional Geometry

The aorta is the largest conduit artery in the human body [1]. In addition to its conduit function, the aorta plays a key role in modulating pulsatile arterial hemodynamics, mediated by its cushioning function of the intermittent left ventricular ejection. Aortic structural parameters (including geometry and wall stiffness) have been shown to be key determinants of aortic hemodynamic function [1, 2]. Despite its prominent age-associated changes and its key hemodynamic role, studies related to structural properties of aortic geometric phenotypes (AGPs) are lacking [3‒5]. For example, previous cross-sectional studies among 210–250 patients undergoing aortic imaging demonstrated that the aorta elongates with age and that thoracic aortic length is greater among patients with acute aortic dissection [3, 4]. Moreover, other studies have been limited to two-dimensional cross-sectional geometric parameters of the aorta, neglecting important three-dimensional (3D) aspects of its geometry, including elongation, tortuosity/unfolding, and curvature, all of which influence aortic function.

3D tomographic imaging (i.e., magnetic resonance imaging, computed tomography [CT]) has become a valuable resource for quantifying structural properties of the heart. Previous analyses of cardiac phenotypes acquired through the segmentation of cardiac magnetic resonance (CMR) imaging data have developed a comprehensive atlas of cardiac structure and function in over 50,000 participants of the UK Biobank (UKB) [6, 7]; however, such analyses of 3D aortic geometry have not been performed to date.

In this study, we present a deep learning approach to segment and comprehensively characterize the 3D geometry of the thoracic aorta, and measure key phenotypes (diameter, length, curvature, tortuosity/unfolding) across various thoracic aortic subsegments. Our segmentation approach consists of (1) modality-specific image segmentation and (2) 3D mesh phenotype extraction. We deployed our segmentation approach on imaging data from 54,241 participants enrolled in the UKB [8] (UKB) and 8,456 individuals enrolled in the Penn Medicine Biobank (PMBB) [9].

The UKB is composed of >500,000 participating individuals aged 37–73 years at the time of recruitment, who underwent various questionnaires, physical measurements, biological sampling (blood and urine), and genome sequencing across 22 assessment centers in the UK [10]. A subset of participants were invited to complete an additional examination that included magnetic resonance imaging of the heart [8]. Data from 54,241 participants enrolled in the UKB who underwent CMR imaging were included in this analysis [8] (Table 1).

The PMBB is composed of >250,000 consenting patients of the Penn Medicine health network [9]. Additionally, all participants medical records including imaging results are de-identified and linked to their identifier. We identified 8,456 patients enrolled who had thoracic CT imaging data (Table 1).

Our method consists of four steps, namely, (1) aortic segmentation, (2) 3D aortic mesh generation, (3) extraction of aortic subsegments, and (4) extraction of AGPs.

We created a dataset of 233 axial steady-state free precision CMR images from the UKB to train and validate a deep learning segmentation algorithm to delineate the thoracic aorta. All CMR images were manually segmented by two trained medical doctors using in-house software. A variation of the U-Net convolutional neural network segmentation architecture was used to segment the aorta [11]. The U-Net architecture is an encoder-decoder consisting of two-dimensional convolutional layers of increasing depth (shown in Fig. 1). The encoder consisted of sequential convolutional and max-pooling layers, allowing the network to learn feature representations across multiple spatial representations. The decoder consisted of an equivalent implementation with max-pooling layers being supplemented for up-sampling layers. Furthermore, all encoder feature representations were concatenated onto their corresponding decoder feature representations. The model was trained on 194 manually segmented CMR scans and validated on an independent subset of 39 CMR scans.

The network was trained using the dice similarity coefficient (DSC) as the loss function and was optimized using the Adam optimizer with a learning rate of 0.001 [12]. Training was performed for 20 epochs with a batch size of 16. Image preprocessing consisted of zero-padding the 238 × 238 axial images to 240 × 240 followed by geometric and intensity augmentations being randomly performed on the training data, and Z-score normalization. Axial CMR image slices were segmented individually before being merged to form the 3D aorta segmentation. The model achieved an average DSC of 0.934 (0.013) on the validation dataset. CT scan data from the PMBB were segmented using the previously developed TotalSegmentator algorithm, achieving a DSC of 0.981 [13]. To extract the thoracic aorta region, the segmentation was cut at the T12 vertebral level.

The voxelized aorta segmentation was postprocessed using morphological erosion and dilation operations with a kernel size of 3 voxels to remove any holes in the segmentation. The processed voxel segmentation was then converted into a 3D aortic mesh using the marching cubes algorithm implemented in the Vascular Modeling Toolkit [14]. Iterative smoothing of the 3D aortic mesh was performed using the Taubin algorithm [15]. Following mesh generation and smoothing, skeletonization was performed to extract the medial axis of the aorta, providing a simplified representation of the vascular structure. Finally, the centerline of the aortic mesh was then interpolated to have 100 evenly spaced points using B-spline interpolation.

Identification of aortic subsegments was accomplished through exploiting properties of the vertical axis (z-axis) of the aorta’s centerline to identify: (1) the apex of the aorta, defined as the maximum centerline point on the vertical axis; (2) the centerline point on the descending aorta with smallest Euclidean distance on the vertical axis with the aortic root. The segment extending from the root to the apex of the aorta was identified as a single segment, hereby named ascending aorta/proximal arch. Similarly, the segment from the aortic apex to the point on the descending aorta vertically corresponding to the level of the aortic root was identified as a single segment, hereby named distal arch/proximal descending aorta. We note that these do not necessarily correspond to standard anatomical segmental definitions, since the latter are defined by the location of the major arch branches, which have variable relationships with the 3D aortic apex [3]. Figure 2a provides an AGP key to understand each region, whereas Figure 2b provides a 3D mesh segmentation.

Aortic curvature was calculated using the Vascular Modeling Toolkit [14]. The aortic radius was computed across every point on the centerline by calculating the nearest distance from the centerline to the aorta mesh and doubled to compute short-axis aortic diameter.

We automatically segmented the entire thoracic aorta and derived 20 unique 3D AGPs from 54,241 UKB participants and 8,456 PMBB participants (Table 2). Representative thoracic aortic meshes are presented for the UKB and PMBB in Figure 3.

This study leveraged aortic imaging data from two large independent cohorts, the UKB and the PMBB, to comprehensively characterize 3D structural parameters of the thoracic aorta. We developed, for the first time, a phenotype extraction protocol consisting of (1) a deep learning segmentation architecture that accurately segments the thoracic aorta, and (2) morphological operations that operate directly on 3D mesh representations to compute aortic length, diameter, tortuosity, curvature and arch measurements, as well as identify subsegments of the aorta. We then successfully deployed our segmentation approach on 54,241 UKB imaging participants and 8,456 PMBB participants (Table 2). Our method provides reliable and quantitative phenotyping of 3D aortic phenotypes. The segmentation approach allows for AGPs to be computed for any aortic segmentation map regardless of its initial modality. This method provides a fully automated approach toward quantifying the 3D structural parameters of the aorta and expands the available phenotypes in two large representative biobanks. This will allow large-scale studies to elucidate the genetics, biological mechanisms, and clinical consequences of aortic degeneration related to aging and disease states in humans.

Previous investigations of 3D AGPs have been limited to small studies requiring extensive manual annotation [3, 5, 16]. These manual approaches, while capable of computing a range of 3D phenotypes such as length, curvature, and tortuosity, are prone to inter- and intraobserver variability. Conversely, some studies using the UKB have explored automated methods for delineating aortic diameter [17]. However, these automated approaches have focused on cross-sectional diameter measurements obtained from transverse aortic cine images at the level of the pulmonary trunk or right pulmonary artery, limiting their scope. Our method combines the comprehensive phenotyping capabilities of manual 3D segmentation with the reliability and scalability of automated approaches, enabling robust and detailed analysis at an unprecedented scale. Additionally, the use of 3D aortic meshes provides a versatile platform for extracting additional phenotypes from any region of the thoracic aorta.

Our method is limited by the image resolution of our tomographic imaging data. Specifically, we were unable to identify the brachiocephalic and left subclavian artery to biologically delineate the aortic arch; however, we were able to provide geometric alternatives that approximate the anatomical regions in question. We note that the aortic subsegments do not necessarily correspond to standard anatomical segmental definitions or aortic subsegments with different embryological origins but were rather derived based on objective unequivocal landmarks in the three-dimensionally segmented thoracic aorta. In particular, the segment extending from the root to the apex of the aorta (segment 1 in Fig. 2a), comprising the proximal transverse aortic arch, was identified as a single segment. Similarly, the segment from the apex to the plane intersecting the aortic root (segment 2 in Fig. 2a) was identified as a single segment, whereas the sum of the two segments above was used to calculate the height and width of segments 1 + 2. Given that this height computation incorporates the ascending aorta and the proximal descending aorta, we note that these height and width indices from our study are not directly comparable to previous studies that measured aortic arch height and width starting from an arbitrarily defined plane intersecting some point of the middle of the anatomic ascending and descending aorta [18].

In conclusion, we present a novel segmentation approach that accurately quantifies the 3D geometry of the thoracic aorta and its key subsegments and provides comprehensive phenotyping of key 3D aortic geometric properties. We leverage our segmentation approach to perform inferencing on all participants in the UKB and PMBB with available imaging data. Future work should investigate the prognostic value as well as the biological foundation of AGPs.

This research has been conducted using the UK Biobank Resource under Application No. 81032. We acknowledge the Penn Medicine BioBank (PMBB) for providing data and thank the patients – participants of Penn Medicine who consented to participate in this research program. We would also like to thank the Penn Medicine BioBank team and Regeneron Genetics Center for providing genetic variant data for analysis. The PMBB is approved under IRB protocol# 813913 and supported by Perelman School of Medicine at University of Pennsylvania, a gift from the Smilow family, and the National Center for Advancing Translational Sciences of the National Institutes of Health under CTSA award number UL1TR001878.

This research has been conducted using the UK Biobank Resource under Application Number 81032. The PMBB is approved under IRB protocol# 813913. Written informed consent was obtained for all participants in this study.

Dr. Chirinos has recently consulted for Bayer, Fukuda-Denshi, Bristol-Myers Squibb, Biohaven Pharmaceuticals, Johnson & Johnson, Edwards Life Sciences, Merck, and NGM Biopharmaceuticals. He received University of Pennsylvania research grants from National Institutes of Health, Fukuda-Denshi, Bristol-Myers Squibb, Microsoft and Abbott. He is named as an inventor in a University of Pennsylvania patent for the use of inorganic nitrates/nitrites for the treatment of Heart Failure and Preserved Ejection Fraction and for the use of biomarkers in heart failure with preserved ejection fraction. He has received payments for editorial roles from the American Heart Association, the American College of Cardiology, Elsevier and Wiley, and payments for academic roles from the University of Texas, Boston University, and Virginia Commonwealth University. He has received research device loans from AtCor Medical, Fukuda-Denshi, Unex, Uscom, NDD Medical Technologies, Microsoft, and MicroVision Medical. The remaining authors have nothing to disclose.

J.A.C. is supported by NIH Grants R01-HL 121510, R33-HL-146390, R01HL153646, R01-AG058969, 1R01-HL104106, P01-HL094307, R03-HL146874, K24-AG070459, and R56-HL136730. W.R.W. is supported by NIH Grants P41-EB029460, R01 HL169378, R01 HL137984, and UL1 TR001878. J.G is supported by R01EB031722 and R01HL133889.

C.B.: writing, methodology, and data analysis. M.J.D.: writing and data analysis. B.Z.: review and interpretation. J.D.A.: data preparation. H.T.: writing and review. H.M.: data preparation. J.D.: data analysis. J.G.: data acquisition and funding. O.S.: data preparation. W.R.W.: data acquisition, supervision, and funding. J.A.C.: conceptualization, supervision, writing, and funding.

Penn Medicine BioBank Banner Author List and Contribution Statements. PMBB leadership team: Daniel J. Rader, M.D., Marylyn D. Ritchie, Ph.D. – contribution: all authors contributed to securing funding, study design, and oversight and all authors reviewed the final version of the manuscript. Patient recruitment and regulatory oversight: JoEllen Weaver, Nawar Naseer, Ph.D., M.P.H., Giorgio Sirugo, M.D., Ph.D., Afiya Poindexter, Yi-An Ko, Ph.D., Kyle P. Nerz – contributions: J.W. managed patient recruitment and regulatory oversight of study. N.N. managed participant engagement and assisted with regulatory oversight and researcher access. G.S. assisted with researcher access. A.P., Y.K., K.P.N. performed recruitment and enrollment of study participants. Laboratory operations: JoEllen Weaver, Meghan Livingstone, Fred Vadivieso, Stephanie DerOhannessian, Teo Tran, Julia Stephanowski, Salma Santos, Ned Haubein, Ph.D., Joseph Dunn – contributions: J.W., M.L., F.V., and S.D. conducted oversight of laboratory operations. M.L., F.V., A.K., S.D., T.T., J.S., and S.S. performed sample processing. N.H. and J.D. are responsible for sample tracking and the laboratory information management system. Clinical informatics: Anurag Verma, Ph.D., Colleen Morse Kripke, M.S. DPT, MSA, Marjorie Risman, M.S., Renae Judy, B.S., Colin Wollack, M.S. – contributions: all authors contributed to the development and validation of clinical phenotypes used to identify study subjects and (when applicable) controls. Genome informatics: Anurag Verma Ph.D., Shefali S. Verma, Ph.D., Scott Damrauer, M.D., Yuki Bradford, M.S., Scott Dudek, M.S., Theodore Drivas, M.D., Ph.D. – contributions: A.V., S.S.V., and S.D. were responsible for the analysis, design, and infrastructure needed to quality control genotype and exome data. Y.B. performed the analysis. T.D. and A.V. provided variant and gene annotations and their functional interpretation of variants.

UK Biobank data are available to researchers following UKB application approval and IRB approval. Researchers can apply at https://www.ukbiobank.ac.uk/. Access to Penn Medicine BioBank data is provided to investigators at the University of Pennsylvania. External access can be provided through collaboration with an investigator at the university. Access to any code used in this analysis will be made available on reasonable request. Further inquiries can be directed to the corresponding author. A preprint version of this article is available on bioRxiv [19].