Segmentation of Glomeruli Within Trichrome Images Using Deep Learning

Shruti Kannan, Laura Morgan, Benjamin Liang, McKenzie G. Cheung, Christopher Q. Lin, Dan Mun, Ralph Nader, Mostafa Belghasem, Joel Henderson, Jean Francis, Vipul C. Chitalia, Vijaya B. KolachalamabioRxiv 2018 · DOI

TL;DR

This paper introduces a deep learning framework for automatically identifying and segmenting glomeruli in trichrome-stained kidney biopsy images. By leveraging a pre-trained Convolutional Neural Network (CNN), the system achieves high accuracy in classifying different glomeruli types and segmenting globally sclerosed glomeruli, offering a potential solution to the laborious manual methods currently used in renal pathology.

What Problem Does This Paper Address?

Accurate assessment of glomeruli, the filtering units of the kidney, is crucial for diagnosing and managing kidney diseases. This assessment, particularly the quantification of normal, partially sclerosed, and globally sclerosed glomeruli, is a standard component of renal pathology reports. However, current methods rely on manual evaluation by pathologists, which is time-consuming, labor-intensive, and prone to inter-observer variability, leading to a lack of standardization. This manual process is particularly challenging given the increasing volume of kidney biopsies and the need for efficient, objective analysis. The paper aims to address this by developing an automated deep learning approach for glomerular segmentation.

Methods

The core of the methodology involves a deep learning approach using a Convolutional Neural Network (CNN). Specifically, the researchers employed a pre-trained Inception V3 architecture, a well-established CNN model for image classification. This model was fine-tuned on a custom dataset of kidney biopsy images.

Key methodological steps include:

  • Image Cropping: A sliding window operation was used to systematically crop large, digitized trichrome-stained kidney biopsy images into smaller, manageable patches. This allowed for focused analysis of individual glomeruli or non-glomerular regions.
  • Expert Annotation: These cropped images were meticulously labeled by three expert pathologists into three categories: "No glomerulus," "Normal or partially sclerosed glomerulus (NPS)," and "Globally sclerosed glomerulus (GS)."
  • CNN Training: The fine-tuned Inception V3 CNN was trained using these labeled cropped images to classify them into the three categories.
  • Data Augmentation: To improve model robustness and generalization, techniques like random whitening were applied to the training data. This process introduced controlled noise and variability into the images, helping to prevent overfitting.
  • Segmentation Pipeline: For the segmentation of GS glomeruli, a pipeline was developed. This involved using the trained CNN to identify regions likely containing GS glomeruli, generating heatmaps to indicate confidence levels, and then applying image processing techniques such as Otsu binarization, distance transform, and watershed segmentation to precisely delineate the boundaries of the identified GS glomeruli.

Datasets

The study utilized a dataset of 275 trichrome-stained images from renal biopsies of 171 chronic kidney disease patients treated at the Boston Medical Center between 2009 and 2012.

  • Training/Testing Split: The dataset was split at the patient level to ensure that images from the same patient were not present in both the training and testing sets. This was done to accurately assess the model's performance on unseen data and account for intra- and inter-patient variability. The split ratio was 70% for training and 30% for testing.
  • Cropped Image Generation: The sliding window operation resulted in a total of 745 images identified as containing glomeruli (611 NPS and 134 GS) and 751 images representing non-glomerular regions. These cropped images formed the basis for model training.

Key Insights and Ideas

  • Transfer Learning Efficacy: The study highlights the power of transfer learning, using a pre-trained CNN (Inception V3) and fine-tuning it on a relatively smaller, specialized dataset. This approach significantly reduces the need for massive datasets typically required for training deep learning models from scratch.
  • Data Augmentation for Robustness: Techniques like random whitening proved effective in improving the CNN model's performance and generalizability, demonstrating their value in enhancing model robustness against variations in image quality and content.
  • Multi-Stage Segmentation Pipeline: The successful segmentation of GS glomeruli was achieved not just by the CNN's classification but by a carefully designed post-processing pipeline that leveraged heatmaps and established image processing algorithms. This demonstrates a practical approach to translating classification outputs into precise segmentation masks.
  • Potential for Automation in Pathology: The work strongly suggests that deep learning can automate complex tasks in digital pathology, potentially freeing up pathologists' time for more nuanced diagnostic work and improving the consistency of assessments.

Key Results

  • Glomerular Classification: The CNN model achieved high accuracy in discriminating between non-glomerular images and those containing NPS or GS glomeruli. On the test data, the model achieved an accuracy of 92.67±2.02% and a Kappa score of 0.8681±0.0392.
  • Glomerular Segmentation: The segmentation model, built upon the best-performing classification model, demonstrated good performance in marking GS glomeruli. The model achieved a sensitivity of 0.558, an F1-score of 0.623, and a Matthews correlation coefficient (MCC) of 0.628 for segmenting GS glomeruli. Model specificity for non-glomerular regions was very high at 0.999.

Limitations and Future Work

  • Limited Scope of Segmentation: The current model was primarily designed to segment globally sclerosed (GS) glomeruli. It could distinguish between "No glomerulus" and "GS glomerulus" but did not explicitly segment "Normal or partially sclerosed (NPS)" glomeruli.
  • Manual "Gold Standard": The classification of cropped images into the three categories relied on manual annotation by experts. The authors acknowledge that a more robust "gold standard" for glomerular identification could further minimize bias.
  • Future Directions: The researchers plan to develop a model capable of distinguishing between normal and partially sclerosed glomeruli, which would provide a more comprehensive assessment of glomerular health. They also emphasize the need for further validation of the framework across different clinical practices and image datasets to ensure its broad applicability. Extending the approach to other staining protocols is also a promising avenue.

Summary updated

Abstract

ABSTRACT Introduction The number of glomeruli and glomerulosclerosis evaluated on kidney biopsy slides constitute as standard components of a renal pathology report. Prevailing methods for glomerular assessment remain manual, labor intensive and non-standardized. We developed a deep learning framework to accurately identify and segment glomeruli from digitized images of human kidney biopsies. Methods Trichrome-stained images (n=275) from renal biopsies of 171 chronic kidney disease patients treated at the Boston Medical Center from 2009-12 were analyzed. A sliding window operation was defined to crop each original image to smaller images. Each cropped image was then evaluated by three experts into three categories: (a) No glomerulus, (b) Normal or partially sclerosed glomerulus and (c) Globally sclerosed glomerulus. This led to identification of 751 unique images representing non­glomerular regions, 611 images with either normal or partially sclerosed (NPS) glomeruli and 134 images with globally sclerosed (GS) glomeruli. A convolutional neural network (CNN) was trained with cropped images as inputs and corresponding labels as output. Using this model, an image processing routine was developed to scan the test data images to segment the GS glomeruli. Results The CNN model was able to accurately discriminate non-glomerular images from NPS and GS images (Performance on test data - Accuracy: 92.67±2.02% and Kappa: 0.8681±0.0392). The segmentation model that was based on the CNN multi-label classifier accurately marked the GS glomeruli on the test data (Matthews correlation coefficient = 0.628). Conclusion This work demonstrates the power of deep learning for assessing complex histologic structures from digitized human kidney biopsies.
BibTeX
@misc{kannan2018segmentation,
  author = {Shruti Kannan and Laura Morgan and Benjamin Liang and McKenzie G. Cheung and Christopher Q. Lin and Dan Mun and Ralph Nader and Mostafa Belghasem and Joel Henderson and Jean Francis and Vipul C. Chitalia and Vijaya B. Kolachalama},
  title = {{Segmentation of Glomeruli Within Trichrome Images Using Deep Learning}},
  year = {2018},
  doi = {10.1101/345579},
  howpublished = {bioRxiv},
}
Figure