VoxelMorph: A Learning Framework for Deformable Medical Image Registration.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Traditional medical image registration is computationally intensive, limiting its application in large-scale studies.
• Existing learning-based methods offer potential but require further development for efficiency and accuracy.

Purpose of the Study:

• Introduce VoxelMorph, a novel learning-based framework for fast, deformable medical image registration.
• Compare unsupervised and semi-supervised training strategies for VoxelMorph.
• Evaluate VoxelMorph's speed and accuracy against state-of-the-art methods.

Main Methods:

• VoxelMorph utilizes a convolutional neural network (CNN) to learn a direct mapping from image pairs to deformation fields.
• Two training strategies were explored: unsupervised learning using image intensity matching and semi-supervised learning leveraging auxiliary segmentation data.
• The framework was trained on a dataset and evaluated for registration accuracy and speed.

Main Results:

• The unsupervised VoxelMorph model achieved accuracy comparable to state-of-the-art methods while being orders of magnitude faster.
• VoxelMorph trained with auxiliary data demonstrated improved registration accuracy.
• Registration accuracy was shown to be influenced by training set size.

Conclusions:

• VoxelMorph offers a significant speedup for medical image registration, making it suitable for large datasets and complex pipelines.
• The framework facilitates advancements in learning-based medical image analysis and its applications.
• VoxelMorph's open-source availability promotes further research and development in the field.