Robust Quad-Tree based Registration on Whole Slide Images

Christian Marzahl, Frauke Wilm, Franz F. Dressler, Lars Tharun, Sven Perner, Christof A. Bertram, Christine Kröger, Jörn Voigt, Robert Klopfleisch, Andreas K. Maier, Marc Aubreville, Katharina BreiningerCOMPAY @ MICCAI 2021 · OpenReview

Abstract

Whole Slide Images (WSIs) in digital pathology present significant registration challenges due to their gigapixel scale, tissue deformations, and diverse sample types. We present a multi-scale, quad-tree-based registration algorithm that recursively decomposes images into patches for piece-wise affine transformation estimation. Validated across four distinct scenarios—multi-scanner tissue, multi-scanner cytology, in-situ proteomic, and multi-slide multi-scanner—the method achieves sub-pixel accuracy (~2 microns). Notably, it provides a robust solution for cytology slides where traditional methods often fail due to the lack of contiguous tissue.

Introduction

In digital pathology, registering Whole Slide Images (WSIs) is essential for multi-modal analysis, such as comparing different scanners or chemical stains (e.g., H&E and IHC). However, the massive size of these images and the non-linear deformations inherent in tissue preparation make standard registration computationally prohibitive. Existing tools often rely on heavy downsampling or tissue segmentation masks, which lose critical cellular detail or fail on cytology samples where cells are sparsely distributed.

Methodology

The proposed method utilizes a Quad-Tree (QT) backbone to implement a "divide and conquer" registration strategy.

  1. Recursive Decomposition: Starting at a low-resolution "Level 0," the algorithm recursively splits the image into quadrants if registration accuracy is insufficient, typically reaching Level 2 (16 patches).
  2. Keypoint Matching: Within each quadrant, SIFT (Scale-Invariant Feature Transform) landmarks are identified and matched between source and target images using k-nearest neighbors.
  3. Transformation Estimation: Transformation matrices (handling translation, rotation, and scaling) are calculated for each segment using RANSAC or Coherent Point Drift.
  4. Piece-wise Affine Alignment: By applying individual transformations to each quadrant, the global result approximates complex non-linear deformations through a series of local linear pieces.

Datasets and Experimental Scenarios

The algorithm was validated on 60 WSIs across four distinct scenarios designed to test robustness across hardware and preparation methods:

  • MSC (Multi-scanner Cytology): Equine lung fluid samples.
  • MST (Multi-scanner Tissue): Canine tumor samples.
  • ISC (In-situ Proteomic): Single slides re-stained and re-scanned.
  • MSSC (Multi-slide Multi-scanner): Consecutive physical tissue slices.
Overview of the four experimental registration scenarios: Multi-scanner Cytology (MSC), Multi-scanner Tissue (MST), In-situ Proteomic (ISC), and Multi-slide Multi-scanner (MSSC).

Results and Discussion

The Quad-Tree approach significantly outperformed state-of-the-art methods like Elastix, ANTs, and ReStain.

Accuracy and Robustness

On MST and MSC datasets, the method achieved a Target Registration Error (TRE) of approximately 2 microns. For context, a red blood cell is about 7 microns wide; this alignment is precise enough for cell-level analysis. While traditional methods showed errors in the hundreds of microns, the QT method remained stable.

Cytology Breakthrough

A key contribution is the method's performance on cytology slides. Because it does not require a pre-segmentation of "contiguous tissue," it can align scattered cells across a slide—a task that previously stumped existing tools.

Detailed view of the Multi-scanner Cytology (MSC) registration, demonstrating alignment of scattered cellular structures.

Efficiency

Operating on full-resolution images through the Quad-Tree preserves cellular details that are lost in downsampling-based methods (like those in the ANHIR challenge). Reaching Level 2 of the Quad-Tree took approximately 134 seconds per gigapixel image, offering a high-precision result in a reasonable timeframe.

Conclusion and Future Work

The Quad-Tree registration method provides a robust, high-precision solution for digital pathology. Its ability to handle cytology samples and approximate non-linear deformations through piece-wise affine transformations marks a significant advancement. Future work will address the challenges of consecutive slices (MSSC), where structural variations between layers limit keypoint matching. Potential improvements include deep-learning-based keypoint detectors or intensity-based refinement.

Summary updated

BibTeX
@inproceedings{marzahl2021robust,
  author = {Christian Marzahl and Frauke Wilm and Franz F. Dressler and Lars Tharun and Sven Perner and Christof A. Bertram and Christine Kröger and Jörn Voigt and Robert Klopfleisch and Andreas K. Maier and Marc Aubreville and Katharina Breininger},
  title = {{Robust Quad-Tree based Registration on Whole Slide Images}},
  booktitle = {COMPAY @ MICCAI},
  year = {2021},
  url = {https://openreview.net/forum?id=HIod0gWrIZv},
}
Figure