Combining MRF-based deformable registration and deep binary 3D-CNN descriptors for large lung motion estimation in...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) achieve state-of-the-art results in medical imaging tasks.
• Accurate correspondence finding, particularly in image registration, remains a challenge.
• Current methods often rely on handcrafted feature descriptors.

Purpose of the Study:

• To adapt deep learning advancements for accurate correspondence finding in medical image registration.
• To develop a hybrid approach integrating deep learned features into discrete optimization-based registration.
• To improve the accuracy and robustness of image registration, specifically for non-rigidly deformed lung CT scans.

Main Methods:

• A two-step hybrid approach was proposed.
• Step 1: Training a deep network on a landmark retrieval task to extract binary local descriptors.
• Step 2: Utilizing these descriptors in a Markov Random Field (MRF)-regularized dense displacement sampling for efficient similarity computations during registration.

Main Results:

• CNN-based descriptors excelled in auxiliary keypoint correspondence tasks.
• Self-similarity-based descriptors provided more accurate registration results.
• A combination of CNN-based and handcrafted descriptors generated the most robust features for registration.

Conclusions:

• A 3D framework for large lung motion estimation was presented.
• The framework combines CNN-based and handcrafted descriptors within a discrete registration method.
• Combining learned and handcrafted features is recommended for future research in this area.