Probabilistic 3D Correspondence Prediction from Sparse Unsegmented Images.

Area of Science:

• Medical imaging analysis
• Biomedical engineering
• Computational anatomy

Background:

• Statistical shape modeling (SSM) is vital for analyzing anatomical form and function in clinical research.
• Traditional SSM pipelines are complex and limited by linearity assumptions, hindering the capture of clinically relevant variations.
• Deep learning advances allow direct SSM inference from images, but struggle with poor data quality or sparsity.

Purpose of the Study:

• To propose SPI-CorrNet, a unified model for predicting 3D correspondences from sparse imaging data.
• To address limitations of current deep learning methods for SSM in challenging imaging conditions.
• To quantify aleatoric uncertainty for reliable clinical deployment.

Main Methods:

• SPI-CorrNet utilizes a teacher network for feature learning regularization.
• The model quantifies data-dependent aleatoric uncertainty by predicting intrinsic input variances.
• The approach enables direct inference of SSM from sparse or low-quality medical images.

Main Results:

• SPI-CorrNet demonstrated enhanced accuracy and robustness in generating SSM from sparse imaging data.
• Experiments on LGE MRI left atrium and Abdomen CT-1K liver datasets validated the model's performance.
• The method effectively handles poor imaging data quality and limited information.

Conclusions:

• SPI-CorrNet offers a robust solution for sparse image-driven statistical shape modeling.
• The model improves the reliability of SSM in clinical applications, particularly with challenging imaging data.
• Quantifying aleatoric uncertainty is key for trustworthy AI in medical image analysis.