Fully Bayesian VIB-DeepSSM.

Area of Science:

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

Background:

• Statistical shape modeling (SSM) is crucial for analyzing anatomical variations and aiding clinical diagnosis.
• Deep learning models can predict SSM from 3D images but often lack robust uncertainty quantification.
• Existing Variational Information Bottleneck DeepSSM (VIB-DeepSSM) offers aleatoric uncertainty but not epistemic uncertainty.

Purpose of the Study:

• To develop a fully Bayesian Variational Information Bottleneck (VIB) framework for DeepSSM.
• To implement and evaluate scalable Bayesian approaches, including concrete dropout and batch ensemble, for enhanced uncertainty inference.
• To improve the accuracy and uncertainty calibration of probabilistic shape prediction from medical images.

Main Methods:

• Derivation of a fully Bayesian VIB formulation for DeepSSM.
• Implementation of concrete dropout and batch ensemble methods for scalable epistemic uncertainty estimation.
• Introduction of a novel combination of concrete dropout and batch ensemble for multimodal marginalization and improved uncertainty calibration.

Main Results:

• The fully Bayesian VIB network accurately predicts statistical shape models from 3D images.
• The proposed methods effectively quantify both aleatoric and epistemic uncertainties.
• Experiments on synthetic and real (left atrium) data show improved uncertainty reasoning without compromising predictive accuracy.

Conclusions:

• Fully Bayesian VIB DeepSSM provides a principled approach for accurate and reliable probabilistic shape modeling from images.
• Scalable Bayesian implementations enhance uncertainty quantification, crucial for clinical applications of SSM.
• The novel combination of dropout and ensemble methods offers superior uncertainty calibration.