An information theoretic approach for non-rigid image registration using voxel class probabilities.

Area of Science:

• Medical image analysis
• Computational anatomy
• Biomedical imaging

Background:

• Accurate non-rigid image registration is crucial for medical image analysis, particularly in comparing images across subjects or aligning atlases.
• Existing methods often rely solely on intensity information, which can be insufficient for complex anatomical variations.

Purpose of the Study:

• To introduce novel information-theoretic similarity measures for non-rigid image registration that incorporate tissue class information.
• To enable more accurate image registration by leveraging prior segmentation data or guiding segmentation with registration.

Main Methods:

• Proposed two similarity measures: one using joint class probability distributions and another minimizing conditional entropy between intensities and class labels.
• Derived analytic gradients for a force field, regularized by a viscous fluid model, to drive the registration process.
• Evaluated performance on non-rigid inter-subject registration and atlas-based segmentation of MR brain images.

Main Results:

• Incorporating tissue class information significantly improved the overlap of corresponding tissue classes after non-rigid matching.
• The proposed class-based measures outperformed registration based solely on intensity information (mutual information).
• Demonstrated the effectiveness in both scenarios: when segmentations are available for both images and when only an atlas segmentation is available.

Conclusions:

• The novel information-theoretic measures effectively integrate tissue class information into non-rigid image registration.
• These methods offer a powerful approach for combining segmentation and registration into a unified process.
• The findings open new avenues for improving medical image analysis and segmentation accuracy.