YeaZ: A convolutional neural network for highly accurate, label-free segmentation of yeast microscopy images

Nicola Dietler, Matthias Minder, Vojislav Gligorovski, Augoustina Maria Economou, Denis Alain Henri Lucien Joly, Ahmad Sadeghi, Chun Hei Michael Chan, Mateusz Koziński, Martin Weigert, Anne‐Florence Bitbol, Sahand Jamal RahibioRxiv 2020 · DOI

TL;DR

This paper introduces YeaZ, a U-Net-based convolutional neural network (CNN) for highly accurate, label-free segmentation of yeast microscopy images. YeaZ, along with a comprehensive dataset and user-friendly interface, addresses the bottleneck of image processing in large-scale yeast experiments, enabling new biological discoveries by efficiently analyzing thousands of cells.

What Problem Does This Paper Address?

Analyzing microscopy images of budding yeast (Saccharomyces cerevisiae) is crucial for many biological studies, but it's a significant bottleneck. The process of "segmentation" – identifying individual cell boundaries – is challenging due to cells crowding together, having irregular shapes, and subtle events like budding. Existing methods often require manual corrections, especially in complex experimental conditions like those involving mutants or different cell cycle stages. This manual effort limits the scale of experiments and the depth of biological insights that can be gained.

Methods

The core of the YeaZ system is a U-Net convolutional neural network (CNN) trained for cell segmentation. A key innovation is a "cell-cell boundary test" that helps resolve oversegmentation issues without requiring fluorescent markers to identify cell interiors, a common workaround in other methods. For tracking cells over time, they employ a bipartite graph matching algorithm. The CNN itself assigns a score to each pixel indicating its likelihood of being part of a cell. This score is then processed through a series of steps: initial thresholding, distance transform to find cell centers, watershed segmentation to delineate regions, and finally, the boundary test to correct erroneous splits.

Datasets

The authors introduce a new, publicly available dataset of over 10,000 high-quality segmented yeast images. This dataset is diverse, including images of wild-type cells, mutants, stressed cells, and time-course experiments. They also utilized existing benchmark datasets for comparison.

Key Insights and Ideas

  • Label-free segmentation is powerful: By avoiding the need for fluorescent markers, YeaZ simplifies experimental setups and allows researchers to use all available fluorescent channels for other measurements.
  • A diverse dataset is key: The success of YeaZ hinges on its training on a large and varied dataset, enabling it to generalize well to different cell shapes and conditions.
  • Iterative improvement is feasible: The authors demonstrate that adding new examples of challenging cell shapes (like filamentous mutants) to the training set can significantly improve the CNN's performance, suggesting a path for continuous improvement.
  • Efficient segmentation unlocks new biology: The ability to process thousands of cells quickly and accurately allows for the detection of subtle biological phenomena, such as the early and gradual nature of morphogenesis control in yeast.

Key Results

  • YeaZ achieves highly accurate segmentation, outperforming existing methods like YeastSpotter and Wood et al. on benchmark datasets.
  • Error rates for YeaZ are exceptionally low, with some datasets showing near-zero errors (e.g., 0.25% for dataset 9, 0% for dataset 10).
  • The system successfully identifies subtle features like tiny buds early in their development.
  • When applied to analyze cell geometry in wild-type and cyclin mutant cells, YeaZ revealed that morphogenesis control occurs earlier and more gradually than previously understood.

Limitations and Future Work

The authors acknowledge that while their dataset is diverse, there might be conditions under which the CNN performs poorly. They suggest that adding more challenging, semi-automatically segmented training examples can improve performance. They also propose an iterative approach of segmenting, retraining, and correcting as a labor-efficient way to expand the CNN's capabilities. The techniques developed are also suggested to be potentially useful for segmenting images of other organisms.

Summary updated

Abstract

Abstract The processing of microscopy images constitutes a bottleneck for large-scale experiments. A critical step is the establishment of cell borders (‘segmentation’), which is required for a range of applications such as growth or fluorescent reporter measurements. For the model organism budding yeast ( Saccharomyces cerevisiae ), a number of methods for segmentation exist. However, in experiments involving multiple cell cycles, stress, or various mutants, cells crowd or exhibit irregular visible features, which necessitate frequent manual corrections. Furthermore, budding events are visually subtle but important to detect. Convolutional neural networks (CNNs) have been successfully employed for a range of image processing applications. They require large, diverse training sets. Here, we present i) the first set of publicly available, high-quality segmented yeast images (>10’000 cells) including mutants, stressed cells, and time courses, ii) a corresponding U-Net-based CNN, iii) a Python-based graphical user interface (GUI) to efficiently use the system, and iv) a web application to test it ( www.quantsysbio.com ). A key feature is a cell-cell boundary test which avoids the need for additional input from fluorescent channels. A bipartite graph matching algorithm tracks cells in time with high reliability. Our network is highly accurate and outperforms existing methods on benchmark images recorded by others, suggesting it transfers well to other conditions. Furthermore, new buds are detected early with high reliability. We apply the system to detect differences in geometry between wild-type and cyclin mutant cells. Our results indicate that morphogenesis control occurs unexpectedly early in the cell cycle and is gradual, demonstrating how the efficient processing of large numbers of cells uncovers new biology. Our system can serve as a resource to the community, expanded continuously with new images. Furthermore, the techniques we develop here are likely to be useful for other organisms as well. The identification of cell borders (‘segmentation’) in microscopy images constitutes a bottleneck for large-scale experiments. For the model organism Saccharomyces cerevisiae , current segmentation methods face challenges when cells bud, crowd, or exhibit irregular features. Here, we present i) the first set of publicly available, high-quality segmented yeast images (>10’000 cells) including mutants, stressed cells, and time courses, ii) a corresponding convolutional neural network (CNN), iii) a graphical user interface and a web application ( www.quantsysbio.com ) to efficiently employ, test, and expand the system. A key feature is a cell-cell boundary test which avoids the need for fluorescent markers. Our CNN is highly accurate, including for buds, and outperforms existing methods on benchmark images, indicating it transfers well to other conditions. To demonstrate how efficient, large-scale image processing uncovers new biology, we analyzed the geometries of ≈2200 wild-type and cyclin mutant cells and found that morphogenesis control occurs unexpectedly early and gradually.
BibTeX
@misc{dietler2020yeaz,
  author = {Nicola Dietler and Matthias Minder and Vojislav Gligorovski and Augoustina Maria Economou and Denis Alain Henri Lucien Joly and Ahmad Sadeghi and Chun Hei Michael Chan and Mateusz Koziński and Martin Weigert and Anne‐Florence Bitbol and Sahand Jamal Rahi},
  title = {{YeaZ: A convolutional neural network for highly accurate, label-free segmentation of yeast microscopy images}},
  year = {2020},
  doi = {10.1101/2020.05.11.082594},
  howpublished = {bioRxiv},
}
Figure