Learning deep abdominal CT registration through adaptive loss weighting and synthetic data generation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deformable image registration is crucial for medical image analysis.
• Convolutional neural networks (CNNs) show promise for image registration.
• Improving CNN-based registration for abdominal imaging requires optimized training strategies.

Purpose of the Study:

• To explore and enhance training strategies for CNN-based image-to-image deformable registration.
• To improve the accuracy and efficiency of abdominal image registration using deep learning.

Main Methods:

• Investigated various training strategies, loss functions, and transfer learning.
• Introduced an augmentation layer for on-the-fly synthetic data generation.
• Implemented a loss layer for dynamic loss weighting.

Main Results:

• Training with segmentation guidance significantly improved deep learning-based registration.
• Transfer learning from brain MRI to abdominal CT datasets enhanced performance.
• Dynamic loss weighting provided marginal improvements without affecting inference speed.

Conclusions:

• Developed a framework (DDMR) that enhances VoxelMorph performance using simple concepts.
• Demonstrated the effectiveness of transfer learning and segmentation guidance for abdominal image registration.
• Further validation on diverse datasets is recommended for the DDMR framework.