Unsupervised adaptive sampling graph autoencoder for 3D surface encoding and mesh representation transfer

Area of Science:

• Computer Vision
• Medical Image Analysis
• Geometric Deep Learning

Background:

• Standardizing anatomical surfaces is crucial for population-level modeling and medical image analysis.
• Existing mesh autoencoders often yield uninterpretable encodings and require restrictive assumptions, limiting accuracy and generalizability.

Purpose of the Study:

• To introduce the Adaptive Sampling Graph Autoencoder (ASGAE), an unsupervised framework for standardized anatomical surface representations.
• To decouple surface geometry from mesh topology for interpretable and pose-standardized encodings.
• To enable mesh representation transfer between heterogeneous surfaces.

Main Methods:

• ASGAE utilizes a novel Fuzzy Segmentation module for probabilistic mapping of diverse input surfaces.
• It learns a geometrically consistent latent subspace in a fully unsupervised manner.
• A novel Adaptive Sampling Decoder reconstructs surfaces, allowing for mesh topology transfer.

Main Results:

• ASGAE achieved state-of-the-art performance in encoding and reconstruction on synthetic and real-world datasets.
• Demonstrated low point-to-surface coordinate (0.43 ± 0.07 mm) and texture reconstruction errors (4.61 ± 0.87%).
• Successfully enabled mesh representation transfer with minimal error (0.44 ± 0.08 mm) and showed utility in landmark identification and pathology classification.

Conclusions:

• ASGAE provides interpretable and pose-standardized surface encodings, overcoming limitations of existing methods.
• The framework enables mesh standardization and representation transfer, critical for medical image analysis.
• ASGAE demonstrates significant potential for advancing population-level modeling and downstream analysis tasks in medical imaging.