Deformation equivariant cross-modality image synthesis with paired non-aligned training data

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Cross-modality image synthesis is crucial for medical applications.
• Existing methods struggle with paired but misaligned data.
• A robust solution for real-world datasets is needed.

Purpose of the Study:

• To develop a generic solution for cross-modality image synthesis with misaligned paired data.
• To introduce novel deformation equivariance encouraging loss functions.
• To enable robust adversarial training on challenging datasets.

Main Methods:

• Joint training of image synthesis and registration networks.
• Utilizing new deformation equivariance encouraging loss functions.
• Conditional adversarial training with misaligned input data.

Main Results:

• A generic and robust method for cross-modality image synthesis with misaligned data.
• Successful adversarial training even with non-aligned image pairs.
• Demonstrated applicability to a wider range of real-world datasets.

Conclusions:

• The proposed method significantly improves cross-modality image synthesis for misaligned data.
• It lowers the barrier for developing new clinical applications.
• Enables effortless training for more difficult medical imaging datasets.