Segmentation and classification of two-channel C. elegans nucleus-labeled fluorescence images

Mengdi Zhao, Jie An, Haiwen Li, Jiazhi Zhang, Shang-Tong Li, Xue-Mei Li, Meng-Qiu Dong, Heng Mao, Louis TaoBMC Bioinformatics 2017 · OpenReview

TL;DR

This paper presents an automated image processing pipeline for analyzing C. elegans nuclei from fluorescence microscopy images. By fusing two-channel fluorescence data and employing a novel segmentation approach, they achieve high accuracy in identifying and classifying nuclei, paving the way for more efficient and objective studies of aging and cellular morphology.

What Problem Does This Paper Address?

Aging is a complex biological process characterized by the breakdown of cellular structures, with nuclear abnormalities being a key indicator, particularly in conditions like progeria. Studying these age-related changes in the nucleus of C. elegans, a widely used model organism, is crucial for understanding aging. However, existing methods for analyzing these nuclei from fluorescence images are often insufficient. The challenges include:

  • Fuzzy boundaries and clumped nuclei: Nuclei in C. elegans images can have indistinct edges and often appear clustered together, making them difficult to isolate.
  • Image quality variations: The images exhibit a wide range of nuclear sizes, high background noise, and variations in image quality, further complicating segmentation.
  • Manual analysis limitations: Biologists traditionally rely on manual identification and subjective assessment of nuclear morphology, which is time-consuming, inefficient, and lacks standardized criteria.

This paper aims to bridge this gap by developing an automatic, robust, and efficient image processing method for segmenting and classifying these challenging fluorescence images.

Methods

The authors propose a four-stage pipeline:

  1. Two-channel Image Fusion: To overcome the limitations of individual channels (green for nuclear lamina, red for chromatin), they fuse the two-channel images. This fusion enhances the contrast and clarity of nuclei, compensating for the low intensity and fuzzy boundaries often seen in the green channel.
  2. Segmentation: This is a multi-step process:
  3. Thresholding Segmentation: Initially, Otsu's method is used for global thresholding to roughly segment nuclei. Local thresholding is applied when necessary to handle variations.
  4. Seed-based Segmentation: The binary image is converted into a distance map, and local maxima are identified as "seeds" representing potential nuclei. A novel seed merging strategy is employed to handle cases with multiple seeds per nucleus. Watershed transform is then used to split clustered nuclei.
  5. Precise Segmentation: A k-means clustering algorithm is applied to refine the nucleus boundaries. This involves creating windows around roughly segmented nuclei and clustering pixels based on their intensity and binary image values, leading to more accurate outlines.
  6. Feature Extraction: A comprehensive set of geometric, intensity, and texture features are extracted from the segmented nuclei. This includes features like area, perimeter, circularity, mean intensity, and texture descriptors derived from the Gray Level Co-occurrence Matrix (GLCM).
  7. Feature Selection and Classification:
    • Feature Selection: The Minimum Redundancy Maximum Relevance (mRMR) method is used to select the most informative features, aiming for high relevance to the target class and low redundancy among features.
  8. Classification: Several machine learning algorithms (SVM, Random Forest, k-NN, Decision Tree, Neural Network) are trained and compared using the selected features to classify nuclei into different tissue types (hypodermal, intestinal, muscle, neuronal).

Datasets

The study utilizes a dataset of 1364 groups of two-channel fluorescence images from two C. elegans strains: wild type (MQD1658) and a long-lived mutant (MQD1798, daf-2(e1370)). Images were acquired from worms of various ages (days 1, 4, 6, 10, 12, 14, 16). The dataset was used for both training and evaluating the segmentation and classification models. No new benchmark dataset was explicitly introduced, but the challenges presented by their specific image acquisition and biological context make it a valuable resource for this type of analysis.

Key Insights and Ideas

  • Fusion is Key: Simply using one channel isn't enough. Fusing the green (lamina) and red (chromatin) channels significantly improves nucleus segmentation by providing complementary information and enhancing contrast.
  • Multi-stage Segmentation is Necessary: The complexity of the images (clumped, fuzzy nuclei) requires a multi-stage segmentation approach, moving from rough thresholding to seed-based splitting and finally precise boundary refinement with k-means.
  • Feature Engineering Matters: A combination of geometric, intensity, and texture features is crucial for distinguishing different nucleus types. The selection of the top 5 features (area, ellipticity, curvature mean, solidity, and GLCM homogeneity) highlights their importance.
  • Random Forest Dominance: For this specific dataset and feature set, Random Forest emerged as the superior classifier, achieving very high accuracy. This suggests its robustness in handling the variations present.
  • Age-Related Changes as Biomarkers: The quantitative analysis of hypodermal nuclei revealed distinct age-related changes in area and solidity between wild-type and long-lived daf-2 mutant worms, suggesting these morphological features could serve as indicators of aging and longevity.

Key Results

  • Segmentation Accuracy: The proposed segmentation method achieved a high precision (98.66% on average) and sensitivity (74.35% on average) for segmented nuclei, outperforming single-channel segmentation. Specifically, 88.31% of nuclei were correctly segmented using the two-channel approach, a significant improvement over single-channel methods.
  • Classification Accuracy: Random Forest achieved a mean class accuracy (MCA) of 98.69% for classifying nuclei into four tissue types. Individual class accuracies were also very high, with muscle and neuronal nuclei reaching 100% accuracy.
  • Morphological Analysis: The study identified age-dependent changes in hypodermal nuclei. Wild-type worms showed a bell-shaped trend in nuclear area, peaking around day 10, while daf-2 mutants exhibited smaller nuclei and a slower decrease in solidity over time, suggesting these are potential indicators of daf-2 longevity.

Limitations and Future Work

  • Out-of-Focus Nuclei: While image fusion helps, the paper acknowledges that not all nuclei are perfectly in focus, and this out-of-plane issue could still affect segmentation.
  • 3D Nature of Nuclei: The current analysis is largely based on 2D projections. The authors suggest that future work could incorporate 3D information for a more complete understanding.
  • Additional Image Modalities: The authors note that bright-field images contain valuable information about worm body size and nucleus location, which could be integrated in future studies.
  • Broader Biological Questions: The developed framework is presented as a tool to facilitate quantitative nuclear aging studies, implying its potential application to a wider range of biological questions beyond the specific analyses presented.

Summary updated

BibTeX
@article{zhao2017segmentation,
  author = {Mengdi Zhao and Jie An and Haiwen Li and Jiazhi Zhang and Shang-Tong Li and Xue-Mei Li and Meng-Qiu Dong and Heng Mao and Louis Tao},
  title = {{Segmentation and classification of two-channel C. elegans nucleus-labeled fluorescence images}},
  journal = {BMC Bioinformatics},
  year = {2017},
  url = {https://openreview.net/forum?id=aGGriHa4yOP},
}
Figure