Learning-based deformable image registration: effect of statistical mismatch between train and test images.

Area of Science:

• Medical imaging
• Artificial intelligence
• Computational anatomy

Background:

• Convolutional neural networks (CNNs) show potential for rapid deformable image registration.
• A key challenge is assessing CNN generalization to unseen data with varying statistical properties.

Purpose of the Study:

• To investigate CNN generalization in deformable image registration under statistical data mismatch.
• To evaluate the impact of image statistics (noise, resolution, power spectrum) on registration accuracy.

Main Methods:

• A UNet-based CNN was trained on simulated CT images with varied noise, resolution, and deformation.
• Target registration error was measured against differences in training and testing data statistics.
• Performance was assessed on both simulated and real anatomical data.

Main Results:

• Registration error increased with statistical mismatch, minimized when training and test data matched.
• CNNs trained on diverse statistical data demonstrated robust generalization.
• Simulated data with matched statistics to real data yielded good performance on anatomical images.

Conclusions:

• CNN generalization in image registration is influenced by statistical properties of training/testing data.
• Training with diverse or matched statistics enhances robustness and performance.
• Characterizing statistical mismatch is vital for deploying CNNs in medical imaging.