A machine learning algorithm for creating isotropic 3D aortic segmentations from routine cardiac MR localizers.

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Aortic aneurysm detection and measurement are critical clinical challenges.
• High-resolution 3D cardiac magnetic resonance (CMR) imaging provides detailed aortic assessment but is time-consuming, limiting its use in routine screening and population studies.

Purpose of the Study:

• To develop and validate a method for generating 3D isotropic aortic segmentations from standard, low-resolution 2D CMR localizer images.
• To assess the clinical suitability and accuracy of these segmentations for identifying aortic aneurysms.

Main Methods:

• A 3D U-Net model (U-Net_LR) was trained using simulated anisotropic 2D localizer images paired with clinician-generated 3D isotropic segmentation masks.
• Segmentation quality and accuracy were evaluated on an external dataset (UK Biobank) and compared against high-resolution 3D isotropic images and a U-Net variant (U-Net_HR) trained on high-resolution data.

Main Results:

• 93% of 3D segmentations generated by U-Net_LR were deemed clinically suitable on an external dataset.
• U-Net_LR demonstrated excellent agreement with ground-truth segmentations (mean DICE score of 0.9), comparable to U-Net_HR.
• Diameter measurements from U-Net_LR, U-Net_HR, and clinical observers showed no significant differences across key aortic regions.

Conclusions:

• A new method effectively produces 3D isotropic aortic segmentations from routine 2D CMR localizers.
• This approach shows strong agreement with segmentations from high-resolution 3D data and holds potential as a non-invasive screening tool for aortic aneurysms, eliminating the need for additional specialized sequences.