A seeding-searching-ensemble method for gland segmentation in H&E-stained images

Yizhe Zhang, Lin Yang, John D. MacKenzie, Rageshree Ramachandran, Danny Z. ChenBMC Medical Informatics Decis. Mak 2016 · DOI

Abstract

Automated gland segmentation in Hematoxylin and Eosin (H&E) stained histology images is a critical task in digital pathology, essential for diagnosing conditions like inflammatory bowel disease and prostate cancer. However, the complexity of tissue backgrounds and the presence of "leaky" boundaries often hinder traditional segmentation methods. We present a novel seeding-searching-ensemble framework that combines local graph-based searches with an ensemble-driven boundary probability map. Our approach achieves significant improvements in both detection and segmentation accuracy, outperforming existing state-of-the-art methods.

Introduction

In clinical pathology, the morphology of glands—including their size, shape, and spatial organization—serves as a "gold standard" for disease grading. Automating the analysis of these structures in H&E-stained images is notoriously difficult due to the inherent messiness of histology slides. Challenges include complex backgrounds with diverse cell types, patient-to-patient variability, and faint staining that results in "weak" boundaries. These factors often lead to "segmentation leaks," where algorithms fail to distinguish gland edges from the surrounding stroma, resulting in inaccurate quantitative measurements.

Methodology

The proposed "seeding-searching-ensemble" method utilizes a multi-stage pipeline designed for robustness and speed.

  1. Sampling and Searching: The process begins by placing a uniform grid of "seeds" across the image. For each seed, a local search identifies the most likely gland boundary by transforming the area into polar coordinates and using a directed acyclic graph (DAG) to find the shortest cyclic path.
  2. Ensemble Boundary Probability Map (BPM): Rather than relying on a single search, the method aggregates suggestions from all local searches to construct a BPM. This map filters out noise and bias, representing the collective likelihood of true gland boundaries.
  3. Segmentation Proposals: A watershed algorithm is applied to the BPM to generate non-overlapping region proposals.
  4. Classification and Refinement: Features such as Histograms of Oriented Gradients (HOG) are extracted, and a Random Forest regressor scores each proposal. A final refinement step merges over-segmented regions and eliminates false positives.
Visualization of the gland segmentation process. Panels (a) and (c) show original H&E-stained tissue; (b) and (d) show the resulting segmentations with red overlays.

Experimental Setup

The study utilized a dataset of 10 clinical H&E-stained histology images of human intestinal tissues (colon and ileum) at 40x magnification. The dataset included normal tissue as well as chronic and acute inflammation. A total of 1,723 glands were manually labeled by experts to provide a ground truth for evaluation.

Results

The proposed method was compared against Object-Graphs (OG) and Polar-Space Random Field (PSRF) approaches.

  • Detection Performance: The ensemble method achieved an F-score of 0.878, significantly outperforming PSRF (0.815) and OG (0.676).
  • Segmentation Accuracy: At the pixel level, the method reached an F-score of 0.856, compared to 0.785 for PSRF.

Qualitative analysis confirms that the method is superior at preserving true boundaries and preventing leaks into the stroma compared to previous models.

Discussion and Conclusion

The core innovation of this work is the ensemble strategy, which provides stability against local artifacts that typically cause single-search methods to fail. By employing graph searches in polar space, the framework imposes hard geometric smoothness constraints, preventing unrealistic segmentation results. While highly irregular glands in advanced disease states still present challenges, the seeding-searching-ensemble framework offers a computationally efficient and accurate solution for clinical digital pathology. Future work will focus on integrating deep learning architectures and expanding the dataset to other organ types.

Summary updated

BibTeX
@article{zhang2016seeding,
  author = {Yizhe Zhang and Lin Yang and John D. MacKenzie and Rageshree Ramachandran and Danny Z. Chen},
  title = {{A seeding-searching-ensemble method for gland segmentation in H\&E-stained images}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2016},
  doi = {10.1186/s12911-016-0312-5},
}
Figure