End-to-end affine registration framework for histopathological images with weak annotations

Area of Science:

• Digital pathology
• Biomedical image analysis
• Computer vision

Background:

• Histopathological image registration is crucial for digital pathology.
• Existing deep learning methods struggle with large initial rotations in pathology images.
• Accurate initial alignment is vital for effective image registration.

Purpose of the Study:

• To develop a general framework for end-to-end affine registration of histopathological images.
• To address the challenge of large rotation angles in real-world pathology image pairs.
• To improve the accuracy and efficiency of histopathological image registration.

Main Methods:

• ARoNet framework utilizes Convolutional Neural Networks (CNNs) for feature extraction and fusion.
• Incorporates a rotation recognition network to correct significant rotational misalignments.
• Employs a self-supervised learning task for unsupervised image representation learning.

Main Results:

• ARoNet outperforms existing affine registration algorithms in alignment accuracy, especially with large rotations (e.g., 180 degrees).
• Achieves fast execution time (0.05s per pair) with high registration accuracy and robustness.
• Provides accurate affine initialization for subsequent non-rigid alignments.

Conclusions:

• The ARoNet framework simplifies and accelerates histopathological image registration.
• Demonstrates potential for clinical applications in digital pathology.
• Offers a robust solution for handling challenging image misalignments.