Using image synthesis for multi-channel registration of different image modalities

Area of Science:

• Medical Imaging
• Image Registration
• Computational Anatomy

Background:

• Multi-channel registration is typically limited to images of analogous modalities.
• Registering magnetic resonance (MR) images with different modalities presents a significant challenge.

Purpose of the Study:

• To develop a multi-channel registration approach for MR images with differing modalities.
• To overcome the limitation of analogous modality requirements in multi-channel registration.

Main Methods:

• Utilized random forest regression to synthesize missing modalities from available ones.
• Converted single-channel, multi-modal registration into a two-channel, mono-modal registration.
• Validated the approach using two registration algorithms and five cost functions.

Main Results:

• The proposed multi-channel synthesis approach yielded statistically significant improvements in registration accuracy (p < 0.001).
• Enhancements were observed across both tested registration algorithms and all cost functions.
• Outperformed standard multi-modal registration using mutual information.

Conclusions:

• The random forest-based modality synthesis effectively enables multi-channel registration between different MR image modalities.
• This method offers a robust solution for improving cross-modal image registration accuracy in medical applications.