ANHIR: Automatic Non-Rigid Histological Image Registration Challenge.

Area of Science:

• Biomedical Imaging
• Computational Pathology
• Medical Image Analysis

Background:

• Histological image registration is crucial for analyzing microscopy data.
• Standard algorithms often struggle with the complexity and variability of histology images.
• A standardized challenge is needed to objectively evaluate and advance registration methods.

Purpose of the Study:

• To establish a fair and independent comparison of automatic non-rigid histological image registration algorithms.
• To benchmark existing and novel registration methods on diverse microscopy datasets.
• To identify key characteristics of high-performing registration techniques.

Main Methods:

• The Automatic Non-rigid Histological Image Registration (ANHIR) challenge utilized 8 datasets with 355 images and 18 stains.
• 481 image pairs were registered, with accuracy evaluated using manually placed landmarks.
• Performance of 7 challenge participants and 6 established methods was analyzed.

Main Results:

• The best performing methods employed coarse initial alignment followed by multiresolution non-rigid registration.
• These advanced methods demonstrated superior robustness compared to off-the-shelf solutions.
• Top methods achieved over 98% landmark registration success with a mean accuracy of 0.44% of the image diagonal.

Conclusions:

• Carefully tuned, robust registration strategies significantly improve accuracy on histological images.
• The ANHIR challenge provides valuable insights into state-of-the-art non-rigid registration.
• The challenge datasets and framework remain available for ongoing research and submissions.