Weakly Supervised Bayesian Shape Modeling from Unsegmented Medical Images.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Machine learning for healthcare

Background:

• Statistical Shape Modeling (SSM) is crucial for clinical research, but traditional methods are complex and prone to bias.
• Deep learning has improved SSM prediction from images but relies on fully supervised methods and burdensome training data creation.

Purpose of the Study:

• To develop a weakly supervised deep learning approach for predicting SSM from medical images.
• To reduce the reliance on manual annotation and prior assumptions in SSM construction.

Main Methods:

• Adapted the Bayesian Variational Information Bottleneck DeepSSM (BVIB-DeepSSM) model for weak supervision.
• Utilized point cloud surface representations for supervision instead of ground truth correspondence points.
• Learned shape correspondence in a data-driven manner.

Main Results:

• Achieved accuracy and uncertainty estimation comparable to fully supervised methods.
• Significantly improved the feasibility and reduced the burden of SSM model training.
• Demonstrated a data-driven approach to correspondence learning without prior variability assumptions.

Conclusions:

• Weakly supervised deep learning with point cloud data offers a viable alternative for SSM construction.
• This method enhances the accessibility and practicality of advanced morphometric analyses in clinical research.
• The approach effectively captures anatomical variation without restrictive prior assumptions.