Dual attention network for unsupervised medical image registration based on VoxelMorph

Area of Science:

• Medical Imaging
• Neuroscience
• Machine Learning

Background:

• Accurate medical image registration is vital for neuroscience and clinical research.
• Existing methods face challenges in achieving high accuracy for 3D deformable registration.

Purpose of the Study:

• To introduce DAVoxelMorph, an unsupervised learning network designed to improve 3D deformable medical image registration accuracy.
• To enhance registration performance through novel architectural modifications and regularization techniques.

Main Methods:

• Developed DAVoxelMorph based on the VoxelMorph model, incorporating a dual attention architecture (spatial and coordinate attention).
• Introduced a bending penalty in the loss function to regularize the deformation field.
• Evaluated performance using metrics such as average Dice scores and percentage of locations with non-positive Jacobian.

Main Results:

• DAVoxelMorph achieved a higher average Dice score (0.714) compared to VoxelMorph (0.703), CycleMorph (0.705), ANTs SyN (0.707), and NiftyReg (0.694).
• Demonstrated a lower percentage of locations with non-positive Jacobian (0.345) compared to VoxelMorph (0.355) and NiftyReg (0.549), indicating reduced deformation issues.
• The model showed increased sensitivity and overall registration accuracy.

Conclusions:

• DAVoxelMorph offers improved accuracy and robustness for 3D deformable medical image registration.
• The dual attention mechanism and bending penalty are effective in enhancing registration performance.
• This unsupervised learning approach shows significant potential for neuroscience and clinical applications.