Unsupervised Non-Rigid Histological Image Registration Guided by Keypoint Correspondences Based on Learnable Deep...

Area of Science:

• Digital pathology
• Medical image analysis
• Computer vision

Background:

• Histological image registration is crucial for analysis but challenging due to staining variations.
• Unsupervised deep learning (DL) methods use keypoint correspondences to guide registration.
• Existing methods struggle with large displacements and tissue tears.

Purpose of the Study:

• To propose an iterative keypoint correspondence-guided (IKCG) unsupervised network for non-rigid histological image registration.
• To improve registration accuracy for images with substantial appearance differences and large displacements.
• To develop a method that addresses challenges like tissue tears in histological slices.

Main Methods:

• Introduced fixed and learnable deep features as keypoint descriptors for establishing correspondences.
• Utilized the distance between keypoint correspondences as a loss function for training the registration network.
• Employed an iterative training strategy to jointly train the registration network and optimize learnable deep features.

Main Results:

• The IKCG network effectively handles local non-rigid large displacement problems.
• Fixed deep features from pre-trained DL networks offer discriminative descriptors.
• Learnable deep features from intermediate layers capture unique histological image information.
• The method achieved 1st rank on both ANHIR and ACROBAT benchmarks by August 6th, 2024.

Conclusions:

• The proposed IKCG unsupervised network significantly advances non-rigid histological image registration.
• Joint optimization of learnable features and iterative training enhances robustness to large displacements and tissue damage.
• The method's superior performance is validated by its top ranking on established benchmarks.