A convolutional neural network for segmentation of yeast cells without manual training annotations

Area of Science:

• Cellular dynamics and microscopy
• Computational biology and image analysis
• Machine learning in biological imaging

Background:

• Single-cell time-lapse microscopy generates large datasets for studying cellular processes.
• Automated software for cell identification and tracking is crucial for analyzing microscopy data.
• Convolutional neural networks (CNNs) show promise for image analysis but require extensive manual annotation for training.

Purpose of the Study:

• To develop a novel approach for training CNNs using synthetic data for yeast cell segmentation.
• To create a software tool for generating synthetic brightfield images of budding yeast cells.
• To implement and evaluate a Mask R-CNN model trained on synthetic data for accurate yeast cell segmentation and tracking.

Main Methods:

• Generation of synthetic brightfield images mimicking budding yeast cells using a dedicated software tool.
• Training a Mask R-CNN model on the fully synthetic dataset for cell segmentation.
• Application of a density-based spatial clustering algorithm (DBSCAN) for cell tracking across time-lapse sequences.

Main Results:

• The Mask R-CNN model trained on synthetic data achieved excellent performance in segmenting real yeast microscopy images.
• The synthetic data generation tool successfully bypassed the need for laborious manual annotation.
• The integrated approach enabled accurate cell segmentation and tracking, demonstrating the utility of synthetic data in machine learning for microscopy.

Conclusions:

• Synthetic data generation is a viable and efficient alternative to manual annotation for training CNNs in biological image analysis.
• The developed tools facilitate the creation of more powerful and user-friendly image processing solutions for yeast cell microscopy.
• This approach significantly advances the potential of automated analysis in time-lapse microscopy studies.