Automated Torso Contour Extraction from Clinical Cardiac MR Slices for 3D Torso Reconstruction

Area of Science:

• Medical Imaging
• Computational Cardiology
• Biomedical Engineering

Background:

• Electrocardiogram (ECG) interpretation is crucial for diagnosing cardiac electrical abnormalities.
• Anatomical variability, particularly torso geometry, significantly influences ECG characteristics.
• Current ECG analysis often overlooks patient-specific anatomical differences.

Purpose of the Study:

• To develop a novel pipeline for reconstructing patient-specific torso and cardiac anatomy from cardiac magnetic resonance images.
• To enable personalized ECG simulations that account for individual torso geometry.
• To improve the accuracy of ECG interpretation by incorporating anatomical data.

Main Methods:

• Utilized standard cardiac magnetic resonance (CMR) images for anatomical reconstruction.
• Developed a two-stage deep learning approach for automated torso contour extraction: an initial u-net segmenter followed by a refinement network.
• Employed a two-channel input (original image and initial segmentation) for the refinement network to enhance contour accuracy.

Main Results:

• The refinement network significantly improved torso contour extraction performance on an independent test set.
• Reduced Hausdorff distance from 9.1 to 4.3 pixels.
• Increased Dice coefficient from 0.75 to 0.93, indicating superior segmentation accuracy.

Conclusions:

• The proposed pipeline accurately reconstructs torso and cardiac anatomy for personalized ECG simulations.
• Accounting for patient-specific torso geometry can significantly impact ECG parameters like QRS.
• This method holds potential for developing clinical tools to enhance ECG interpretation by integrating anatomical data.