From Images to Probabilistic Anatomical Shapes: A Deep Variational Bottleneck Approach

Area of Science:

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

Background:

• Statistical shape modeling (SSM) from 3D medical images is valuable for pathology detection and morphology analysis.
• Deep learning simplifies SSM but requires calibrated uncertainty for clinical trust.
• Current methods use isolated PCA shape representations, limiting flexibility and linearity.

Purpose of the Study:

• To develop a principled framework for probabilistic shape prediction from 3D medical images.
• To overcome limitations of existing SSM techniques, particularly regarding uncertainty quantification.
• To enable more scalable, flexible, and non-linear shape modeling.

Main Methods:

• Proposed a framework based on variational information bottleneck theory.
• Learned latent shape representations in the context of the learning task, avoiding supervised encoding.
• Predicted probabilistic shapes directly from 3D images without pre-defined descriptors.

Main Results:

• The proposed model achieved improved accuracy compared to state-of-the-art methods.
• Demonstrated better calibrated aleatoric uncertainty estimates.
• The model is self-regularized and generalizes well with limited data.

Conclusions:

• The novel framework offers a more scalable and flexible approach to SSM from medical images.
• It provides accurate probabilistic shape predictions with reliable uncertainty quantification.
• This advances the clinical applicability of deep learning-based SSM.