Deep Medial Voxels: Learned Medial Axis Approximations for Anatomical Shape Modeling

Area of Science:

• Medical Image Analysis
• Computational Geometry
• Computer-Aided Diagnosis

Background:

• Shape reconstruction from imaging volumes is crucial in medical image analysis.
• Current methods involve segmentation, post-processing, and meshing, which are time-consuming.
• Neural networks offer automated shape reconstruction but are limited in handling topological variations.

Purpose of the Study:

• To introduce a novel semi-implicit shape representation called deep medial voxels.
• To enable accurate shape reconstruction from imaging volumes, preserving topological details.
• To facilitate improved visualization and computer simulations in medical applications.

Main Methods:

• Developed deep medial voxels, a semi-implicit representation learning approach.
• Utilized convolution surfaces for shape reconstruction from the learned medial representation.
• Focused on faithfully approximating the topological skeleton from imaging volumes.

Main Results:

• The deep medial voxels approach accurately reconstructs shapes from imaging volumes.
• The method preserves topological information, addressing limitations of prior neural network methods.
• The technique demonstrates potential for enhanced visualization and computer simulations.

Conclusions:

• Deep medial voxels offer a robust method for shape reconstruction in medical imaging.
• This semi-implicit representation effectively captures topological skeletons.
• The approach shows promise for advancing medical visualization and simulation capabilities.