Three-dimensional lumbar spine generation using variational autoencoder

Area of Science:

• Medical Imaging and Computer-Aided Diagnosis
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Analysis of lumbar spine diseases necessitates extensive three-dimensional (3D) models.
• A significant gap exists in available 3D lumbar spine models for research, particularly for conditions like scoliosis.
• Current data collection for specific lumbar spine pathologies is time-consuming and challenging.

Purpose of the Study:

• To develop an automated method for generating diverse and authentic 3D lumbar spine models.
• To overcome the limitations of data scarcity in lumbar spine disease research.
• To provide a tool for population-based modeling and potential clinical applications.

Main Methods:

• An end-to-end network utilizing a 3D variational autoencoder was designed for random 3D lumbar spine model generation.
• The network incorporates a dual-path encoder with spatial coordinate attention modules and a regularization loss for enhanced reconstruction.
• Gaussian noise layers were integrated into the decoder to improve model authenticity and diversity.

Main Results:

• Experimental validation on entire lumbar spines and individual vertebrae demonstrated promising performance.
• Quantitative metrics included voxel intersection over union (0.588-0.684) and Dice coefficient (0.739-0.811).
• The developed method showed comparable results to statistical shape models (SSM) but without requiring landmarks.

Conclusions:

• The proposed 3D variational autoencoder network effectively generates realistic and diverse 3D lumbar spine models.
• This fully automatic approach offers a significant advantage over landmark-dependent methods like SSM.
• The method holds potential as a valuable clinical tool for population-based lumbar spine modeling and research.