OTMorph: Unsupervised Multi-Domain Abdominal Medical Image Registration Using Neural Optimal Transport

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Machine learning in healthcare

Background:

• Deformable image registration is crucial for analyzing medical images, especially for abdominal diseases like hepatic cancer and lymphoma.
• Multi-domain abdominal images (different modalities/protocols) are vital but misaligned due to factors like patient movement.
• Current deep learning methods struggle with multi-domain registration due to differing image characteristics (contrast, intensity).

Purpose of the Study:

• To propose a novel unsupervised multi-domain image registration framework, OTMorph, addressing the challenges of domain gaps in medical imaging.
• To improve the accuracy and generalizability of deformable registration for multi-domain abdominal images.

Main Methods:

• OTMorph utilizes neural optimal transport to learn a mapping between moving and fixed image domains.
• A transport module estimates a domain-transported volume, aligning data distributions.
• A subsequent registration module uses this transported volume to accurately estimate deformation fields.

Main Results:

• OTMorph demonstrates superior deformable registration performance on multi-modality and multi-parametric abdominal images.
• The domain-transported image effectively alleviates the domain gap between input images.
• The model shows improved generalizability, even on out-of-distribution data.

Conclusions:

• OTMorph offers a robust solution for unsupervised multi-domain deformable image registration in abdominal medical imaging.
• The neural optimal transport approach successfully bridges domain gaps, enhancing registration accuracy and reliability.
• The framework's generalizability suggests broad applicability for diverse medical image registration tasks.