D2C-Morph: Brain regional segmentation based on unsupervised registration network with similarity analysis.

Area of Science:

• Medical image analysis
• Deep learning applications in neuroscience
• Computational anatomy

Background:

• Brain regional segmentation is crucial for medical image analysis, aiding in diagnosis and prognosis.
• Current deep learning segmentation models require accurate spatial alignment of image data.
• Existing methods often necessitate separate registration and segmentation steps, increasing complexity.

Purpose of the Study:

• To propose D2C-Morph, a novel unsupervised deep learning model for joint image registration and segmentation.
• To improve the accuracy and efficiency of brain image processing pipelines.
• To demonstrate the utility of deformation fields from registration for segmentation tasks.

Main Methods:

• Developed a dual-path network emphasizing input features.
• Implemented unsupervised learning for joint registration and segmentation.
• Utilized contrastive learning twice and a correlation layer for enhanced feature map similarity.

Main Results:

• D2C-Morph successfully performs joint registration and segmentation without supervision.
• The dual-path network and contrastive learning enhance feature representation.
• Correlation layer improves decoder performance by increasing feature map similarity.

Conclusions:

• The proposed D2C-Morph model effectively integrates registration and segmentation.
• Unsupervised learning enables simultaneous processing, simplifying pipelines.
• Deformation fields from registration can be leveraged for improved segmentation accuracy.