Related Experiment Video
Updated: Dec 26, 2025
11:16
In vitro Assessment of Aortic Regurgitation Using Four-Dimensional Flow Magnetic Resonance Imaging
Published on: February 25, 2022
3.7K
Fully automated 3D aortic segmentation of 4D flow MRI for hemodynamic analysis using deep learning
Haben Berhane1, Michael Scott2,3, Mohammed Elbaz2,3
1Department of Medical Imaging, Ann & Robert H. Lurie Children's Hospital of Chicago, Chicago, Illinois.
Magnetic Resonance in Medicine
|March 14, 2020
Summary
Deep learning rapidly automates 3D aortic segmentation from 4D-flow MRI, enabling reproducible quantification of aortic flow and dimensions. This AI approach significantly reduces analysis time compared to manual methods.
Area of Science:
- Cardiovascular imaging and diagnostics
- Artificial intelligence in medical imaging
- Medical image analysis
Background:
- Accurate quantification of aortic dimensions and hemodynamics is crucial for diagnosing and managing cardiovascular diseases.
- Traditional manual segmentation of 4D-flow MRI data is time-consuming and prone to interobserver variability.
- Developing automated methods for aortic segmentation can improve efficiency and reproducibility in clinical practice.
Purpose of the Study:
- To develop and validate a deep learning-based approach for automated 3D segmentation of the aorta using 4D-flow MRI.
- To enable fast and reproducible quantification of aortic flow, peak velocity, and dimensions.
- To assess the performance and reproducibility of the automated segmentation compared to manual analysis.
Main Methods:
- A convolutional neural network (CNN) was trained on 4D-flow MRI data from 1018 subjects to generate 3D aortic segmentations.
- Manual segmentations served as the ground truth for training and validation.
- Performance was evaluated using Dice scores, Hausdorff distance, and average symmetrical surface distance. Flow, peak velocity, and dimensions were quantified and compared using Bland-Altman analysis.
Main Results:
- The CNN achieved automated 3D aortic segmentation in under a second, significantly faster than manual analysis (0.438s vs. 630s).
- The deep learning model demonstrated excellent segmentation performance with a median Dice score of 0.951.
- Quantification of aortic flow, peak velocity, and dimensions showed excellent agreement with manual analysis, with limits of agreement <10% for most parameters.
Conclusions:
- Deep learning enables rapid and automated 3D aortic segmentation from 4D-flow MRI, facilitating efficient clinical workflows.
- The automated method provides reproducible quantification of aortic hemodynamics and dimensions.
- Further research should explore the utility of this deep learning approach for other vascular structures and across different imaging vendors.
