Graph convolutional network with probabilistic spatial regression: application to craniofacial landmark detection...

Area of Science:

• Medical imaging
• Computer vision
• Biomedical engineering

Background:

• Accurate quantitative evaluation of pediatric craniofacial anomalies is crucial for diagnosis and treatment.
• Existing landmark detection methods are often unsuitable for unstructured 3D photogrammetry data.
• 3D photogrammetry presents challenges due to variable mesh structures (vertices and polygons).

Purpose of the Study:

• To develop a robust method for anatomical landmark detection in 3D photogrammetry data.
• To overcome limitations of current methods in handling unstructured 3D mesh data.
• To achieve state-of-the-art accuracy in calculating landmark coordinates from 3D photographs.

Main Methods:

• Proposed a graph-based convolutional neural network utilizing Chebyshev polynomials.
• Exploited vertex coordinates, polygonal connectivity, and surface normal vectors for feature extraction.
• Introduced a novel weighting scheme for feature aggregation and a trainable regression scheme for landmark coordinate calculation without vertex correspondence.

Main Results:

• The proposed method successfully extracted multi-resolution spatial features from 3D photographs.
• A novel trainable regression scheme enabled accurate landmark coordinate calculation.
• Achieved state-of-the-art landmark detection errors, demonstrating superior performance.

Conclusions:

• The developed graph-based convolutional neural network offers a significant advancement in 3D craniofacial landmark detection.
• The method's ability to work without vertex correspondence enhances its applicability to diverse 3D photogrammetry datasets.
• This approach holds promise for improving the quantitative evaluation of pediatric craniofacial anomalies.