[Paper Review] Automatic Nonrigid Histological Image Registration with Adaptive Multistep Algorithm

Marek Wodziński, Andrzej Skalski|arXiv (Cornell University)|Apr 1, 2019

Advanced Image and Video Retrieval Techniques8 references4 citations

TL;DR

This paper presents a robust, fully automatic nonrigid histological image registration method for the ANHIR challenge, combining preprocessing, multiple initial alignment strategies, and four parallel nonrigid registration algorithms (local affine, Demons, MIND Demons, and TPS). The method achieved a 0.38% average median rTRE with 99.792% robustness, using an automatic selection strategy to choose the best result from diverse algorithms.

ABSTRACT

In this paper, we present a short description of the method proposed to ANHIR challenge organized jointly with the IEEE ISBI 2019 conference. We propose a method consisting of preprocessing, initial alignment, nonrigid registration algorithms and a method to automatically choose the best result. The method turned out to be robust (99.792% robustness) and accurate (0.38% average median rTRE). The main drawback of the proposed method is relatively high computation time. However, this aspect can be easily improved by cleaning the code and proposing a GPU implementation.

Motivation & Objective

• Address the challenge of automatic, nonrigid registration of multimodal histological images with missing data and intensity variations.
• Develop a robust, fully automatic pipeline that requires no human intervention or parameter tuning.
• Improve registration accuracy across diverse, difficult histological image pairs, especially those with structural mismatches or intensity differences.
• Achieve high performance in the ANHIR challenge, where ranking is based on median rTRE, necessitating minimal error even in hard cases.

Proposed method

• Preprocess images by converting to grayscale, smoothing, downsampling to resolution-specific scales (1024–4096 pixels), histogram matching to balance intensity distributions, zero-padding, and intensity inversion.
• Perform initial alignment using SIFT, ORB, and SURF features with RANSAC to estimate similarity transformations, selecting the best based on Dice coefficient of thresholded masks.
• Apply an alternative initial alignment using Li thresholding and iterative rotation optimization with SSD and affine registration when the first method fails.
• Run four nonrigid registration algorithms in parallel: locally affine registration with brightness/contrast correction and missing data probability estimation, symmetric compositive Demons, MIND Demons using SSD between MIND descriptors, and TPS using keypoint matches from initial alignment.
• Automatically select the best nonrigid result using the lowest MIND SSD between a masked region in the fixed image and the corresponding transformed moving image.
• Use a fail-detection mechanism to identify and handle difficult cases, ensuring robustness across all 481 training pairs.

Experimental results

Research questions

• RQ1 Can a multimodal, fully automatic nonrigid registration pipeline achieve high accuracy and robustness on challenging histological image pairs with missing data and intensity differences?
• RQ2 How effective is an ensemble of diverse nonrigid registration algorithms (local affine, Demons, MIND Demons, TPS) when combined with automated result selection?
• RQ3 To what extent does histogram matching improve initial alignment, and when is it detrimental, particularly for multimodal images?
• RQ4 Can a simple TPS method based on initial feature matches outperform more complex algorithms in hard cases, and under what conditions?
• RQ5 How does the use of multiple algorithms in parallel, followed by automated selection, improve final ranking in a challenge setting where median rTRE is the key metric?

Key findings

• The method achieved an average median rTRE of 0.38% across all training pairs, indicating high registration accuracy.
• Robustness was 99.792%, with only one failure across 481 image pairs, demonstrating strong reliability.
• The local affine registration method was effective in most cases, but the addition of Demons and MIND Demons improved the median rTRE by 0.04% and corrected three hard failures.
• The TPS interpolation method, though computationally cheap, significantly reduced error in the most difficult cases where other algorithms failed.
• The final result selection based on MIND SSD between masked regions was effective in identifying the best-performing algorithm for each pair.
• Despite high computation time due to single-core implementation of the local affine method, the method’s performance was significantly improved by using higher resolutions (up to 4096 pixels) for faster algorithms like Demons and MIND Demons.