A deep-learning approach for direct whole-heart mesh reconstruction

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Deep learning applications

Background:

• Accurate cardiac surface geometry is crucial for clinical applications.
• Current deep learning methods often rely on voxel-wise segmentation, leading to artifacts and topological errors.
• Limitations include disconnected regions, incorrect topology, and staircase artifacts.

Purpose of the Study:

• To develop a novel deep learning approach for direct whole heart surface mesh reconstruction from volumetric medical images (CT and MR).
• To overcome limitations of traditional segmentation-then-reconstruction methods.
• To generate anatomically consistent and high-resolution cardiac geometries.

Main Methods:

• A graph convolutional neural network (GCNN) approach is proposed.
• The GCNN predicts vertex deformations on a pre-defined mesh template.
• Direct mesh prediction from CT and MR image data.

Main Results:

• The method achieves comparable or superior accuracy to existing deep learning techniques for whole heart reconstruction on both CT and MR data.
• Generates whole heart geometries with improved anatomical consistency.
• Produces high-resolution geometries from lower-resolution inputs.
• Demonstrates temporally consistent predictions for cardiac motion from cine sequences.

Conclusions:

• The proposed direct mesh prediction method offers an effective alternative for cardiac geometry reconstruction.
• It enhances anatomical consistency and resolution compared to prior methods.
• Potential for efficient 4D whole heart dynamics construction from medical imaging.