Cell Motility Dynamics: A Novel Segmentation Algorithm to Quantify Multi-Cellular Bright Field Microscopy Images

Assaf Zaritsky, Sari Natan, Judith Horev, Inbal Hecht, Lior Wolf, Eshel Ben‐Jacob, Ilan TsarfatyPLoS 2011 · DOI

Abstract

Confocal microscopy analysis of fluorescence and morphology is becoming the standard tool in cell biology and molecular imaging. Accurate quantification algorithms are required to enhance the understanding of different biological phenomena. We present a novel approach based on image-segmentation of multi-cellular regions in bright field images demonstrating enhanced quantitative analyses and better understanding of cell motility. We present MultiCellSeg, a segmentation algorithm to separate between multi-cellular and background regions for bright field images, which is based on classification of local patches within an image: a cascade of Support Vector Machines (SVMs) is applied using basic image features. Post processing includes additional classification and graph-cut segmentation to reclassify erroneous regions and refine the segmentation. This approach leads to a parameter-free and robust algorithm. Comparison to an alternative algorithm on wound healing assay images demonstrates its superiority. The proposed approach was used to evaluate common cell migration models such as wound healing and scatter assay. It was applied to quantify the acceleration effect of Hepatocyte growth factor/scatter factor (HGF/SF) on healing rate in a time lapse confocal microscopy wound healing assay and demonstrated that the healing rate is linear in both treated and untreated cells, and that HGF/SF accelerates the healing rate by approximately two-fold. A novel fully automated, accurate, zero-parameters method to classify and score scatter-assay images was developed and demonstrated that multi-cellular texture is an excellent descriptor to measure HGF/SF-induced cell scattering. We show that exploitation of textural information from differential interference contrast (DIC) images on the multi-cellular level can prove beneficial for the analyses of wound healing and scatter assays. The proposed approach is generic and can be used alone or alongside traditional fluorescence single-cell processing to perform objective, accurate quantitative analyses for various biological applications.

Introduction

Cell motility is a fundamental process underlying various physiological and pathological conditions, including tissue repair, immune response, and cancer metastasis. Quantifying how cells move and interact within their environment is crucial for biological research. While fluorescence microscopy provides high-contrast images, it often necessitates the use of exogenous dyes that can be phototoxic or alter natural cell behavior. Bright-field imaging techniques, such as Differential Interference Contrast (DIC), offer a non-invasive alternative but present significant challenges for automated analysis due to low contrast, uneven illumination, and the difficulty of distinguishing individual cells within dense clusters.

Traditional segmentation methods often attempt to identify individual cells, a task that frequently fails in "crowded" environments where cell boundaries are obscured. This paper introduces MultiCellSeg, a machine-learning-based framework that shifts the focus from single-cell detection to the segmentation of entire multi-cellular regions. By accurately defining the boundaries of cellular masses, MultiCellSeg provides a robust and objective means to quantify complex assays like wound healing and cell scattering.

Methodology

The MultiCellSeg algorithm employs a two-stage segmentation process designed for robustness and minimal user intervention.

1. Patch-Based Classification

The input image is partitioned into small, overlapping patches (e.g., 20x20 pixels). For each patch, a set of basic image features is extracted, including intensity statistics and gradient histograms. These features are processed by a cascade of Support Vector Machines (SVMs). The SVMs assign a confidence score to each patch, classifying it as either "cell" or "background." This local approach allows the system to capture the characteristic textures of cellular regions even in noisy bright-field data.

2. Global Refinement and Graph-Cut

To address local classification errors—such as small dark spots in the background being misidentified as cells—the algorithm incorporates a global refinement step. A region classifier evaluates the spatial context of the patches. Finally, a graph-cut algorithm is applied to the confidence map to produce smooth, biologically plausible contours. This ensures that the final segmentation accurately represents the continuous mass of the cell colony.

Texture Analysis for Scatter Assays

For scatter assays, where cells detach from a primary colony in response to stimuli, the researchers utilized Local Binary Patterns (LBP). LBP is a powerful texture descriptor that converts the visual appearance of a cell cluster into a quantitative histogram. This allows the system to distinguish between "compact" and "scattered" morphologies with high precision.

Experimental Results

The performance of MultiCellSeg was evaluated using 126 manually annotated images across several datasets, including the TScratch benchmark, melanoma cell lines, and DA3 mouse mammary adenocarcinoma cells.

Superior Accuracy and Robustness

MultiCellSeg consistently outperformed existing tools like TScratch. Across all tested datasets, the algorithm achieved an average accuracy ranging from 91.2% to 96.9%. Notably, it reduced the error rate of misclassified zones by approximately 30% compared to traditional intensity-based methods. Once trained, the algorithm operates without the need for manual parameter tuning, eliminating human bias.

Biological Insights: Wound Healing and HGF/SF

The tool was used to analyze time-lapse DIC images of wound healing assays. The results revealed that the rate of wound closure is remarkably linear over time. Furthermore, MultiCellSeg quantified the biological effect of Hepatocyte Growth Factor/Scatter Factor (HGF/SF), demonstrating that it accelerates the healing rate by approximately two-fold.

Automated Scatter Assay Scoring

The LBP-based texture analysis achieved 100% accuracy in classifying scatter assay images, matching the performance of expert human observers. This demonstrates that multi-cellular texture is an excellent descriptor for measuring HGF/SF-induced cell scattering.

Conclusion

MultiCellSeg represents a significant advancement in the automated analysis of bright-field microscopy. By leveraging textural information at the multi-cellular level rather than relying on individual cell boundaries, the algorithm provides a robust, parameter-free solution for quantifying cell motility. Its ability to produce accurate, objective data from non-invasive imaging makes it a valuable tool for high-throughput biological screening and the study of dynamic cellular processes. Future work will focus on integrating MultiCellSeg into open-source platforms like CellProfiler and expanding its library of pre-trained models for diverse cell lines.

Summary updated

Abstract

Confocal microscopy analysis of fluorescence and morphology is becoming the standard tool in cell biology and molecular imaging. Accurate quantification algorithms are required to enhance the understanding of different biological phenomena. We present a novel approach based on image-segmentation of multi-cellular regions in bright field images demonstrating enhanced quantitative analyses and better understanding of cell motility. We present MultiCellSeg, a segmentation algorithm to separate between multi-cellular and background regions for bright field images, which is based on classification of local patches within an image: a cascade of Support Vector Machines (SVMs) is applied using basic image features. Post processing includes additional classification and graph-cut segmentation to reclassify erroneous regions and refine the segmentation. This approach leads to a parameter-free and robust algorithm. Comparison to an alternative algorithm on wound healing assay images demonstrates its superiority. The proposed approach was used to evaluate common cell migration models such as wound healing and scatter assay. It was applied to quantify the acceleration effect of Hepatocyte growth factor/scatter factor (HGF/SF) on healing rate in a time lapse confocal microscopy wound healing assay and demonstrated that the healing rate is linear in both treated and untreated cells, and that HGF/SF accelerates the healing rate by approximately two-fold. A novel fully automated, accurate, zero-parameters method to classify and score scatter-assay images was developed and demonstrated that multi-cellular texture is an excellent descriptor to measure HGF/SF-induced cell scattering. We show that exploitation of textural information from differential interference contrast (DIC) images on the multi-cellular level can prove beneficial for the analyses of wound healing and scatter assays. The proposed approach is generic and can be used alone or alongside traditional fluorescence single-cell processing to perform objective, accurate quantitative analyses for various biological applications.
BibTeX
@inproceedings{zaritsky2011cell,
  author = {Assaf Zaritsky and Sari Natan and Judith Horev and Inbal Hecht and Lior Wolf and Eshel Ben‐Jacob and Ilan Tsarfaty},
  title = {{Cell Motility Dynamics: A Novel Segmentation Algorithm to Quantify Multi-Cellular Bright Field Microscopy Images}},
  booktitle = {PLoS},
  year = {2011},
  doi = {10.1371/journal.pone.0027593},
}
Figure