Bayesian Fully Convolutional Networks for Brain Image Registration.

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Machine Learning

Background:

• Nonrigid medical image registration aligns images for tasks like data fusion and pathological analysis.
• Current methods prioritize accuracy, often overlooking the uncertainty inherent in registration results.
• Quantifying registration uncertainty is crucial for identifying abnormal source data.

Purpose of the Study:

• To propose a novel Bayesian fully convolutional neural network for nonrigid medical image registration.
• To incorporate geometric uncertainty mapping for assessing the reliability of registration outcomes.
• To enhance network convergence using group normalization within the Bayesian neural network framework.

Main Methods:

• Developed a Bayesian fully convolutional neural network model for nonrigid medical image registration.
• Integrated a geometric uncertainty map to quantify the uncertainty of the registration transformation.
• Employed group normalization to improve the convergence of the Bayesian neural network.

Main Results:

• The proposed method achieved superior registration accuracy compared to existing learning-based approaches.
• Demonstrated comparable anti-folding performance to established methods like fast image registration and VoxelMorph.
• Successfully generated uncertainty maps, providing a confidence interval for registration results.

Conclusions:

• The Bayesian neural network approach effectively addresses nonrigid medical image registration with improved accuracy.
• The generated uncertainty maps offer valuable insights into the reliability of registration, aiding in abnormality detection.
• Group normalization facilitates stable and efficient training of the Bayesian neural network for medical image registration.