Neural implicit heart coordinates: 3D cardiac shape reconstruction from sparse segmentations

Area of Science:

• Medical Imaging
• Computational Anatomy
• Biomedical Engineering

Background:

• Accurate patient-specific cardiac modeling from limited clinical images is challenging.
• Existing neural implicit functions have limitations in cross-subject anatomical consistency.
• Developing standardized anatomical reference frames for the heart is crucial for robust modeling.

Purpose of the Study:

• Introduce Neural Implicit Heart Coordinates (NIHCs) as a standardized implicit coordinate system for the human heart.
• Enable anatomically consistent 3D cardiac reconstruction from sparse 2D segmentations.
• Improve the efficiency and accuracy of patient-specific cardiac modeling.

Main Methods:

• Developed NIHCs based on universal ventricular coordinates.
• Predicted NIHCs directly from sparse 2D cardiac segmentations.
• Decoded NIHCs into dense 3D segmentations and high-resolution meshes.
• Trained the model on a large dataset of 5000 cardiac meshes.

Main Results:

• Achieved high reconstruction accuracy with mean Euclidean surface errors of 2.51 ± 0.33 mm (diseased) and 2.31 ± 0.36 mm (healthy).
• Demonstrated anatomically coherent reconstruction under severe data sparsity and noise.
• Successfully recovered complex cardiac structures like valve planes.
• Reduced inference time from over 60s to 5-15s compared to traditional methods.

Conclusions:

• NIHCs provide a robust and efficient anatomical representation for 3D cardiac reconstruction.
• The method significantly improves patient-specific modeling from minimal input data.
• NIHCs offer a standardized framework for consistent cardiac anatomy mapping across subjects.