Unsupervised labeling of glomerular boundaries using Gabor filters and statistical testing in renal histology

Brandon Ginley, John E. Tomaszewski, Rabi Yacoub, Feng Chen, Pinaki SarderJournal of Medical Imaging 2017 · DOI

TL;DR

This paper introduces an unsupervised computational pipeline for automatically identifying and segmenting glomerular boundaries in renal histology images. By combining Gabor filters for texture analysis with statistical F-testing and distance transforms, the method achieves high accuracy, paving the way for more efficient and precise quantitative analysis of renal pathology.

What Problem Does This Paper Address?

The glomerulus, the kidney's primary filtering unit, is crucial for understanding kidney health. Damage to glomeruli can lead to serious conditions like proteinuria and kidney failure. The current gold standard for assessing this damage involves manual examination of kidney biopsies under a microscope. However, this manual approach is time-consuming, subjective, and can be prone to qualitative or semi-quantitative assessments, limiting the precision of diagnosis and disease trajectory prediction. A significant hurdle in automating this process is the accurate and automatic identification of complex glomerular boundaries within dense histological images. This paper aims to overcome this challenge by developing a computational method that can reliably segment these boundaries without requiring manual annotation.

Methods

The proposed method is a multi-stage computational pipeline designed for unsupervised glomerular boundary segmentation. It begins with preprocessing steps, including image size reduction and grayscale conversion. A crucial component is Gabor filtering, a technique inspired by mammalian visual systems, used to capture textural information. The Gabor filter bank generates multiple filtered images, and a K-means clustering step is applied to these outputs to create an initial Gabor mask representing potential glomerular areas.

Next, a statistical F-test is employed to refine this mask. This test compares the variance of pixels within the Gabor-identified regions against a local window, helping to distinguish true glomerular structures from background noise. This results in a F-test mask.

To handle gaps and ensure smooth boundaries, both the Gabor mask and the F-test mask undergo distance transformations. A spatial weighting map is also generated, which emphasizes pixels closer to the center of potential glomeruli. These three transformed images (Gabor DT, F-test DT, and spatial weighting map) are then averaged and thresholded to produce the final segmented glomerular boundary. The authors highlight that this combination of techniques allows for robust identification of glomerular boundaries across various staining conditions.

Datasets

The study utilized a dataset of 1000 rat renal glomeruli images stained with five different histological stains: Hematoxylin and Eosin (H&E), Periodic Acid-Schiff (PAS), Gömöri's trichrome, Congo red (CR), and Jones silver. The authors also mention using murine tissue for proof-of-concept demonstrations in disease models (specifically Focal Segmental Glomerulosclerosis - FSGS). No new benchmark datasets were introduced; instead, the performance was evaluated against manual annotations of these existing images.

Key Insights and Ideas

  • Texture as a Key Discriminator: The core insight is that the distinct textural contrast between the dense, packed nuclei within glomeruli and the sparser, more linear tubular nuclei outside them is a highly effective feature for segmentation. Gabor filters are adept at capturing these textural differences.
  • Statistical Validation for Robustness: The integration of an F-test provides a statistically grounded method to validate the initial Gabor-based segmentation, significantly reducing false positives and improving boundary accuracy.
  • Unsupervised Approach is Crucial: The development of an unsupervised method is a significant step forward, as it bypasses the need for large, manually annotated datasets, which are costly and time-consuming to create. This makes the method more scalable and practical for clinical applications.
  • Stain Invariance: The pipeline demonstrates robustness across multiple common histological stains, suggesting its broad applicability in routine pathology workflows.
  • Disease Application Potential: The proof-of-concept on FSGS models shows that the method can identify glomeruli even in the presence of pathological changes, hinting at its utility in studying disease progression.

Key Results

The proposed method achieved a mean sensitivity of 0.88, specificity of 0.96, and accuracy of 0.92 across 1000 rat renal glomeruli images. The H&E and trichrome stains showed the highest performance and lowest variance. The method was also shown to accurately estimate glomerular locations, detecting 87% of glomeruli compared to manual examination. The computational pipeline requires approximately 6 to 10 seconds per biopsy-sized image containing 5 to 15 glomeruli. When compared to the Segmental Histogram of Oriented Gradients (S-HOG) method, this approach achieved a higher percentage of glomeruli scoring an F-measure of 0.8 or greater (99.1% vs. 90.1%).

Limitations and Future Work

  • Parameterization: The method relies on several parameters that require tuning, although the authors provide optimized values for rat renal tissue. They note that these parameters may need adjustment for different species or tissue types.
  • K-means Clustering: The K-means clustering step within the Gabor analysis can be time-consuming, especially for larger images.
  • Disease Complexity: While a proof-of-concept for FSGS was demonstrated, the authors acknowledge that diseased renal tissues present a wide range of structural changes (e.g., altered shapes, partial integrity) that could affect segmentation. Future work will involve validating the model on more diverse diseased tissues.
  • Species Differences: Anatomical differences between species (e.g., human vs. rat/mouse glomeruli) will likely require further modifications to the method for human data application.
  • Future Directions: The authors envision this method as a component of a semi-supervised digital pathology workflow, where pathologists can review and correct segmentations. They aim to further develop computational quantification of renal structures to enable faster diagnosis and intervention for proteinuric renal diseases.

Summary updated

Abstract

The glomerulus is the blood filtering unit of the kidney. Each human kidney contains [Formula: see text] glomeruli. Several renal conditions originate from structural damage to glomerular microcompartments, such as proteinuria, the excessive loss of blood proteins into urine. The gold standard for evaluating structural damage in renal pathology is histopathological and immunofluorescence examination of needle biopsies under a light microscope. This method is limited by qualitative or semiquantitative manual scoring approaches to the evaluation of glomerular structural features. Computational quantification of equivalent features promises to improve the precision of glomerular structural analysis. One large obstacle to the computational quantification of renal tissue is the identification of complex glomerular boundaries automatically. To mitigate this issue, we developed a computational pipeline capable of extracting and exactly defining glomerular boundaries. Our method, composed of Gabor filtering, Gaussian blurring, statistical [Formula: see text]-testing, and distance transform, is able to accurately identify glomerular boundaries with mean sensitivity/specificity of [Formula: see text] and accuracy of 0.92, on [Formula: see text] glomeruli images stained with standard renal histological stains. Our method will simplify computational partitioning of glomerular microcompartments hidden within dense textural boundaries. Automatic quantification of glomeruli will streamline structural analysis in clinic and can pioneer real-time diagnoses and interventions for renal care.
BibTeX
@article{ginley2017unsupervised,
  author = {Brandon Ginley and John E. Tomaszewski and Rabi Yacoub and Feng Chen and Pinaki Sarder},
  title = {{Unsupervised labeling of glomerular boundaries using Gabor filters and statistical testing in renal histology}},
  journal = {Journal of Medical Imaging},
  year = {2017},
  doi = {10.1117/1.jmi.4.2.021102},
}
Figure