Automatic landmark detection and mapping for 2D/3D registration with BoneNet.

Area of Science:

• Biomechanical analysis
• Animal imaging
• Deep learning applications

Background:

• 3D musculoskeletal motion analysis is crucial for biological studies.
• Current methods like image matching and manual annotation have limitations (computational cost, operator dependency).
• Automated landmark detection is needed to overcome these challenges.

Purpose of the Study:

• To develop an automated method for 3D animal pose estimation from X-ray fluoroscopy.
• To address the limitations of existing image matching and manual annotation techniques.
• To introduce a novel deep learning approach for accurate 2D landmark detection.

Main Methods:

• Proposed a two-part strategy: automated 3D landmark extraction and a deep neural network (BoneNet) for 2D landmark detection.
• Utilized shortest voxel coordinate variance for 3D landmark extraction from tomographic reconstructions.
• Developed BoneNet, a customized ResNet18-based neural network, for precise 2D landmark identification in X-ray images.
• Reconstructed 3D animal poses by aligning 2D and 3D landmarks with a reference model.

Main Results:

• BoneNet accurately detected 2D landmarks in simulated, noisy 2D X-ray images.
• The method achieved promising estimations for rigid and articulated parameters.
• Validation was performed on simulated X-ray images from a piglet hindlimb CT scan.
• The approach successfully eliminated the need for manual landmark annotation.

Conclusions:

• The proposed automated method, integrating 3D landmark extraction and BoneNet, effectively estimates 3D animal poses from X-ray data.
• BoneNet demonstrates high accuracy in 2D landmark detection, even with image noise.
• This technique offers a robust and efficient alternative for musculoskeletal motion analysis in biological research.