Learning-based three-dimensional registration with weak bounding box supervision

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Image registration is crucial for medical image analysis.
• Deep learning methods are increasingly used but often require extensive annotations.
• Current deformable image registration methods often rely on conventional algorithms or detailed segmentations.

Purpose of the Study:

• To develop a weakly supervised learning scheme for deformable image registration.
• To train a network for large displacement deformations using only bounding box labels.
• To reduce the annotation effort required for training medical image registration models.

Main Methods:

• Proposed a novel weakly supervised learning scheme for deformable image registration.
• Utilized bounding box labels to calculate the loss function, avoiding densely labeled images.
• Evaluated the model on 3D abdominal CT and MRI interpatient images.

Main Results:

• Achieved significant performance improvements compared to unsupervised methods.
• Demonstrated a performance increase of $$ for CT and 20% for MRI images.
• Outperformed weakly supervised methods that use detailed image segmentations, considering reduced annotation effort.

Conclusions:

• The proposed method enhances image registration performance with minimal annotation effort.
• Weakly supervised learning with bounding boxes is effective for deformable image registration.
• This approach offers a practical solution for training accurate medical image registration models efficiently.