Unsupervised Histological Image Registration Using Structural Feature Guided Convolutional Neural Network.

Area of Science:

• Histological image analysis
• Computational pathology
• Medical image registration

Background:

• Accurate registration of multiple stained histological images is crucial for analysis.
• Supervised methods require laborious ground-truth data, motivating unsupervised approaches.
• Histological image registration faces challenges like staining variance, repetitive textures, and missing tissue sections.

Purpose of the Study:

• To develop an unsupervised method for histological image registration that overcomes limitations of existing approaches.
• To address challenges including appearance variance, repetitive textures, and missing tissue sections in histological images.

Main Methods:

• Proposed an unsupervised Structural Feature Guided convolutional neural network (SFG).
• SFG utilizes structural features robust to staining variations.
• Employs a multi-scale strategy combining low-resolution (rough) and high-resolution (fine) structural features.
• Incorporates dense and sparse structural consistency constraints for local and global information, respectively.
• Dense component uses whole-image structural feature maps; sparse component uses matched key point distances.

Main Results:

• The proposed SFG method demonstrated robustness against multiple staining, repetitive textures, and missing tissue sections.
• Achieved state-of-the-art performance, ranking first on the public histological dataset (ANHIR) as of January 18th, 2022.

Conclusions:

• The unsupervised SFG method offers a powerful solution for histological image registration.
• Structural feature guidance effectively handles common challenges in histological image analysis.
• The method alleviates the need for manual annotation, reducing labor and time costs.