SparseMorph: A weakly-supervised lightweight sparse transformer for mono- and multi-modal deformable image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Deformable image registration (DIR) is vital for clinical diagnosis precision.
• Transformer-based DIR methods excel at long-range dependencies but suffer from high computational costs.
• Enhancing DIR's computational efficiency and accuracy is a key research challenge.

Purpose of the Study:

• To develop a computationally efficient and accurate deformable image registration model.
• To introduce SparseMorph, a lightweight Transformer-based DIR method.
• To improve clinical diagnosis through enhanced image registration.

Main Methods:

• Proposed SparseMorph, a weakly-supervised lightweight Transformer model for DIR.
• Introduced a sparse multi-head self-attention (SMHA) mechanism to reduce computational complexity.
• Developed a multi-branch multi-layer perception (MMLP) module for efficient feature accumulation.
• Implemented an anatomically-constrained, weakly-supervised strategy for region-of-interest alignment.

Main Results:

• SparseMorph outperformed the state-of-the-art TransMatch on mono-modal brain datasets (IXI, OASIS) with improved DSC scores for MRI-to-CT registration.
• Achieved significant DSC score improvements on the multi-modal cardiac dataset (MMWHS) for both MRI-to-CT and CT-to-MRI tasks.
• SparseMorph utilized only 33.33% of TransMatch's parameters, demonstrating superior computational efficiency.

Conclusions:

• SparseMorph offers an efficient and effective solution for both mono- and multi-modal deformable image registration.
• The proposed model surpasses current state-of-the-art algorithms in registration accuracy and computational performance.
• SparseMorph presents a promising DIR method for clinical applications, enhancing diagnostic precision.