Deep Adaptive Log-Demons: Diffeomorphic Image Registration with Very Large Deformations

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Image registration is crucial for analyzing medical images, but handling large deformations is challenging.
• Traditional methods often fail due to inaccuracies in initial affine transformations, especially with significant image biases.
• Existing preregistration techniques can misalign images, negatively impacting subsequent nonrigid registration steps.

Purpose of the Study:

• To develop a robust framework for capturing large and complex deformations in 2D and 3D image registration.
• To improve the accuracy and efficiency of the preregistration step in image registration pipelines.
• To introduce novel similarity metrics for enhanced registration performance.

Main Methods:

• A two-layer deep adaptive registration framework using multilayer convolutional neural networks (CNNs) for 2D image rotation classification and parameter identification.
• A triplanar 2D CNN approach for locating the affine matrix in 3D images via feature correspondences.
• Iterative deformation removal combining preregistration and Demons registration.
• Integration of Principal Component Analysis (PCA) with Pearson and Spearman correlation for new similarity standards.

Main Results:

• The proposed framework achieves more accurate registration results compared to state-of-the-art methods on both synthetic and real datasets.
• The method demonstrates faster convergence speeds in experiments.
• Accurate classification of rotation parameters and separate identification of scale and translation improve preregistration accuracy.

Conclusions:

• The novel framework effectively addresses the limitations of traditional image registration, particularly in handling large deformations.
• The deep learning-based approach offers superior accuracy and efficiency for 2D and 3D image registration.
• The new similarity metrics contribute to improved registration performance and faster convergence.