Weakly-supervised learning of multi-modal features for regularised iterative descent in 3D image registration.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning for medical image registration is advancing but faces challenges with large parameter spaces and limited annotated data.
• Progress in registration lags behind segmentation due to annotation scarcity and complexity.
• Iterative solutions could significantly benefit medical image registration, especially for multi-modal scenarios.

Purpose of the Study:

• To develop a weakly-supervised deep learning approach for disentangling appearance-based feature learning and deformation estimation in medical image registration.
• To improve multi-modal image registration quality, particularly with limited training data.

Main Methods:

• An end-to-end trainable, weakly-supervised deep learning framework for feature extraction.
• Mapping complex appearance features into a common space for registration.
• Evaluation on thoracoabdominal CT and MRI datasets.

Main Results:

• The proposed method demonstrates competitive performance against state-of-the-art hand-crafted features, Mutual Information-based methods, and fully-integrated CNNs.
• The approach effectively handles limitations of small and weakly-labeled training datasets.
• Favorable comparisons in multi-modal registration tasks.

Conclusions:

• Disentangling feature learning and deformation estimation is a promising strategy for improving deep learning-based medical image registration.
• Weakly-supervised learning offers a viable solution for registration tasks with scarce annotations.
• The developed method shows potential for advancing multi-modal medical image registration accuracy and efficiency.