🫀 Ascending Aorta Segmentation Pipeline

This repository is part of our study published in the European Heart Journal – Cardiovascular Imaging: Fay et al., 2025

• 3D segmentation of the thoracic aorta from non-contrast-enhanced MRA data.
• Shape measurement of diameter (at 200 points along the thoracic aorta), length and volume

---

📌 Inference Overview

---

Step 1: Preprocessing

• Rescale and reorient input NIfTI images based on parameters in config/config.yaml:
  • Target spacing
  • Target shape
  • Target orientation
• Normalize image intensities.
• Store preprocessed images in Zarr and HDF5 formats.

---

Step 2: Predict Ascending Aortic Segmentation & Landmark Points

• Load preprocessed images from HDF5 format.
• Process data in overlapping chunks to reduce memory load.
• Use a trained LandmarkNet model to predict:
  • Ascending aorta segmentation
  • Heatmaps for 6 anatomical landmarks
• Reconstruct full-resolution predictions from patch-wise inference.
• Save results to HDF5.

📍 Landmark Points Predicted

No.	Anatomical Landmark	Description
1	Sinotubular junction	Border between aortic root and ascending aorta
2	Brachiocephalic artery origin	Center of vessel branch at aortic wall
3	Left carotid artery origin	Center of vessel branch at aortic wall
4	Left subclavian artery origin	Center of vessel branch at aortic wall
5	Diaphragm	At the level of the hiatus
6	End of image	Lower boundary of the input image

---

Step 3: Postprocessing

• Landmark Detection: Refine heatmap predictions using Gaussian fitting to get sub-voxel landmark positions.
• Mask Postprocessing: Keep only the largest connected component to reduce noise.
• Coordinate Transformation: Convert image coordinates to physical (world) coordinates using affine matrix.

📊 Output Files

File	Format	Content
Results/postprocessed_masks.zarr	Zarr	Segmentation masks with affine metadata
Results/postprocessed_landmarks.json	JSON	Physical coordinates of all detected landmarks

---

Step 4: Shape Measurement

• Centerline Detection: Use the fast marching method (FMM) between landmarks to extract the vessel centerline.
• Centerline Analysis:
  • Normalize by arc length
  • Fit cubic splines
  • Extract tangent vectors and local frames
• Cross-Section Analysis:
  • Generate orthogonal planes
  • Measure area, diameter, shape
  • Project contours to local frames
• Data Storage: All outputs are stored in Zarr format:
  • y: Centerline coordinates
  • tangent: Direction vectors
  • distance_map, travel_time_map: Profiles
  • cs_area, cs_equivalent_diameter, vtk_area
  • landmarks: projected landmark positions

📊 Output Files

File	Format	Content
Results/shape_measurements.zarr	Zarr	Centerline geometry and cross-sectional metrics
Results/shape_measurements.log	Log	Processing logs and duration info

---

📖 Citation

If you use this pipeline, please cite the following paper:

@article{fay2025determinants,
  title={Determinants of ascending aortic morphology: cross-sectional deep learning-based analysis on 25 073 non-contrast-enhanced NAKO MRI studies},
  author={Fay, Louisa and Hepp, Tobias and Winkelmann, Moritz T and Peters, Annette and Heier, Margit and Niendorf, Thoralf and Pischon, Tobias and Endemann, Beate and Schulz-Menger, Jeanette and Krist, Lilian and others},
  journal={European Heart Journal-Cardiovascular Imaging},
  volume={26},
  number={5},
  pages={895--907},
  year={2025},
  publisher={Oxford University Press UK}
}