Low-Dimensional Non-Rigid Image Registration Using Statistical Deformation Models From Semi-Supervised Training Data.

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Machine Learning in Radiology

Background:

• Accurate non-rigid image registration is crucial for medical image analysis, often requiring high-dimensional transformations.
• Statistical Deformation Models (SDMs) aim to simplify these transformations but are challenging to train with limited annotated data.

Purpose of the Study:

• To develop and evaluate a semi-supervised training (SST) framework for learning high-dimensional SDMs.
• To improve the accuracy and robustness of non-rigid image registration using both labeled and unlabeled imaging data.

Main Methods:

• Implemented an SST framework to train non-rigid SDMs, combining 39 labeled MR datasets with over 1200 unlabeled MRIs.
• Applied the trained SDM to inter-subject registration of skull-stripped brain MR images.
• Validated the approach using leave-one-out cross-validation.

Main Results:

• The SST approach significantly improved registration accuracy compared to standard intensity-based methods.
• Achieved a 99% reduction in transformation dimensionality, demonstrating efficient learning of deformation patterns.
• The resulting registration algorithm was robust and accurate.

Conclusions:

• Semi-supervised training is effective for learning complex, high-dimensional statistical deformation models from limited annotated data.
• This framework offers a robust and accurate solution for non-rigid medical image registration, particularly for inter-subject brain MRI analysis.
• The method substantially reduces computational complexity while enhancing performance.