Learning continuous shape priors from sparse data with neural implicit functions

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Machine Learning

Background:

• Statistical shape models (SSMs) are vital for medical image analysis tasks like reconstruction and classification.
• Current SSMs are limited by the resolution of training data, hindering high-resolution shape prior learning from sparse scans.
• Anisotropic scans with large slice distances are common in clinical practice (e.g., CT, MRI), posing challenges for existing methods.

Purpose of the Study:

• To develop a novel shape modeling approach capable of training on sparse, low-resolution medical image data.
• To enable the reconstruction of high-resolution 3D shapes from limited input, overcoming limitations of current methods.
• To create a robust latent space representation for sparse shapes, invariant to acquisition parameters and capable of discriminating between healthy and pathological cases.

Main Methods:

• Utilized neural implicit functions for continuous shape representation.
• Trained the model on sparse, binary segmentation masks with large inter-slice distances.
• Developed a method to embed diverse sparse segmentation masks into a unified, low-dimensional latent space.

Main Results:

• Successfully reconstructed high-resolution shapes from as few as three orthogonal slices.
• Demonstrated the model's ability to create a latent space invariant to acquisition direction, resolution, and spacing.
• Showcased the latent representation's effectiveness in discriminating between healthy and pathological shapes from sparse data.
• Validated the model on lumbar vertebra and distal femur datasets, confirming smooth latent space and characteristic shape variation capture.

Conclusions:

• The proposed neural implicit function-based shape modeling approach effectively addresses the challenge of sparse medical imaging data.
• This method enables high-resolution shape reconstruction and robust feature representation from limited, clinically relevant scans.
• The developed model holds significant potential for improving shape analysis and diagnosis in medical imaging applications.