Leveraging Descriptor Learning and Functional Map-based Shape Matching for Automatic Landmark Acquisition

Area of Science:

• Biological sciences
• Geometric morphometrics
• Computational anatomy

Background:

• Manual landmark placement in geometric morphometrics is time-consuming and limits scalability for large datasets.
• Existing methods require predefined hypotheses, lacking flexibility for theoretical adjustments.
• Automated landmark placement is crucial for efficient analysis of biological shape variation.

Purpose of the Study:

• To investigate the precision and accuracy of landmarks derived from functional correspondences using a deep functional map network.
• To develop and assess an automated landmarking method for biological specimens.
• To compare the proposed method's performance against a state-of-the-art technique (MALPACA).

Main Methods:

• Utilized a deep functional map network to learn shape descriptors and establish point-to-point correspondences between specimens.
• Interrogated functional maps to identify corresponding landmarks based on initial manual placements.
• Applied the automated landmarking process to a dataset of rodent mandibles for comparative analysis.

Main Results:

• The proposed deep functional map-based method demonstrated notable speed improvements over MALPACA while maintaining competitive accuracy.
• Root Mean Square Error (RMSE) analysis showed comparable performance to MALPACA, particularly with smaller training datasets, indicating strong generalizability.
• Visual evaluations confirmed the precision of the automated landmark placements, with acceptable deviations.

Conclusions:

• Unsupervised learning models show significant potential for automating anatomical landmark placement.
• The developed method offers a viable, efficient, and flexible alternative to traditional manual and semi-automated landmarking techniques.
• This approach enhances the scalability and applicability of geometric morphometrics in biological research.