A robust and scalable image analysis framework for phenotyping human thoracic aorta

Q: What is the clinical evidence from this study?

Population: Thoracic aorta phenotyping (n=660). Intervention: Automated nnU-Net and PyTorch-FEA image analysis framework vs. Consensus manual reference annotation. Primary outcome: Automated morphological and biomechanical phenotype extraction rate and landmark localization accuracy.

Q: What does this research mean for the field?

An automated, scalable CT image analysis framework accurately extracts anatomical landmarks, structured quadrilateral meshes, and comprehensive morphological and biomechanical phenotypes of the human thoracic aorta without manual intervention. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Jiasong Chen; Linchen Qian; Ruonan Gong; Christina Sun; Thuy Pham; Catilin Martin; Mohammad Zafar; John Elefteriades; Wei Sun; Liang Liang

doi:10.64898/2026.10.11.758402

October 13, 2026Open Access

A robust and scalable image analysis framework for phenotyping human thoracic aorta

View Full Paper

Ask AI

Bookmark

Share

Key result

Automated CT framework extracts comprehensive aortic valve phenotypes with ~96% success rate.

n=660

Why the study?

Large-scale phenotyping of the thoracic aorta from contrast-enhanced CT requires a workflow in which every step from segmentation to phenotype extraction runs without manual intervention.

Does an automated nnU-Net and PyTorch-FEA image analysis framework accurately extract anatomical landmarks, patient-specific quadrilateral meshes, wall-stress fields, and regional phenotypes of the thoracic aorta from contrast-enhanced CT scans compared to manual reference annotations?

Key Points

To develop and evaluate a fully automated, scalable framework that extracts anatomical landmarks, structured quadrilateral meshes, wall-stress fields, and regional phenotypes of the human thoracic aorta from contrast-enhanced CT scans.
Implemented four nnU-Net segmentation tasks to delineate the aorta, its regions, and aortic root partitions, deriving landmarks and branch ostia from the geometric interfaces between segmented labels.
Fitted a structured template of 24,882 quadrilateral elements to the aortic surface via staged, differentiable mesh deformation and computed transmural mean wall stress using PyTorch-FEA.
Evaluated the automated pipeline on test patients and extracted morphological and biomechanical phenotypes across a cohort of 660 patients.
In two example test patients, the mean distance between automated and reference landmarks was 0.31 mm for cusp nadirs and 0.44 mm for commissures, remaining well below the 1.0 mm image spacing.
Fitted meshes achieved mean distances to the target surface of 0.016 mm and 0.013 mm (maxima of 0.26 mm and 0.32 mm), and finite element wall-stress simulations converged without retries.
The framework successfully processed 631 of 660 patients, generating 191 morphological and 57 biomechanical phenotypes per patient without manual intervention.

Structured PICO

Does an automated nnU-Net and PyTorch-FEA image analysis framework accurately extract anatomical landmarks, patient-specific quadrilateral meshes, wall-stress fields, and regional phenotypes of the thoracic aorta from contrast-enhanced CT scans compared to manual reference annotations?

P

Population

660 patients from the public INSPECT dataset with contrast-enhanced thoracic CT scans, tricuspid aortic valves, and two- or three-branch aortic arches (526 training and 134 test cases).

E

Exposure

Automated image analysis framework using four 3D full-resolution nnU-Net segmentation tasks, a surface-curve-point landmark extraction procedure, staged differentiable mesh deformation of a structured quadrilateral template (24,882 elements), and finite element analysis (PyTorch-FEA) at a uniform static pressure of 16 kPa (120 mmHg).

C

Comparator

Reference annotation established by consensus of three annotators in 3D Slicer, and theoretical Laplace's law wall-stress estimates.

O

Outcome

Accuracy of automated anatomical landmark extraction (distance in mm between automated and reference points for cusp nadirs and commissures), mesh fitting accuracy (point-to-surface distance and element interior angle deviation), wall-stress computation convergence and consistency with Laplace's law, and automated extraction of regional morphological and biomechanical phenotypes.surrogate

This fully automated deep learning and finite-element framework enables scalable, noninvasive extraction of 3D morphological and biomechanical wall-stress phenotypes of the thoracic aorta from standard contrast-enhanced CT scans without manual intervention.

Limitations

Preliminary results for landmark accuracy, remeshing, and wall stress were reported in detail for only two example test patients.
Excluded four-branch aortic arches, non-tricuspid aortic valves, and scans without contrast enhancement.
CT does not provide patient-specific material properties of the aortic wall, requiring a uniform 2 mm wall thickness and stiff neo-Hookean material assumption.
CT does not provide the material properties of the patient's aortic wall, requiring the wall to be modeled as a stiff neo-Hookean material with uniform 2 mm thickness.
Scans without contrast enhancement, poor image quality, non-tricuspid aortic valves, and four-branch arches were excluded.
Detailed landmark, remeshing, and wall-stress accuracy results are reported as preliminary results for only two example patients from the test set.

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Enables scalable aortic phenotyping from routine CT without manual input; leaves open prospective validation for clinical risk stratification.

Abstract

Large-scale phenotyping of the thoracic aorta requires a workflow in which every step, from segmentation to phenotype extraction, runs without manual intervention. We present a robust and scalable framework that converts contrast-enhanced CT into anatomical landmarks, a patient-specific quadrilateral mesh, a wall-stress field and regional phenotypes of the thoracic aorta. Four nnU-Net segmentation tasks delineate the aorta, its anatomical regions and two partitions of the aortic root. A surface–curve–point procedure derives the annulus, the sinotubular junction, the ostia of the supra-aortic branches, the cusp nadirs and the commissures from the interfaces between segmented regions. We fit a structured template of 24,882 quadrilateral elements to the segmented aortic surface by staged, differentiable mesh deformation and compute the transmural mean wall stress with PyTorch-FEA. The structured mesh provides node-to-node correspondence across patients, anchored to the anatomy by the landmarks, from which the framework extracts regional morphological and biomechanical phenotypes.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Chen et al. (2026) studied Thoracic aorta phenotyping (n=660). Automated nnU-Net and PyTorch-FEA image analysis framework vs. Consensus manual reference annotation was evaluated on Automated morphological and biomechanical phenotype extraction rate and landmark localization accuracy. The automated CT framework extracted 191 morphological and 57 biomechanical phenotypes in 631 of 660 patients, with mean errors of 0.31 mm for cusp nadirs and 0.44 mm for commissures in two test cases.

synapsesocial.com/papers/6aee6b6871a759480622093b https://doi.org/10.64898/2026.10.11.758402

View Full Paper

Ask AI

Bookmark

Share

A robust and scalable image analysis framework for phenotyping human thoracic aorta

Key Points

Structured PICO

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

A robust and scalable image analysis framework for phenotyping human thoracic aorta

Key Points

Structured PICO

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading