Cell-APP: A generalizable method for cell annotation and cell-segmentation model training.

Area of Science:

• Cellular biology
• Computer vision
• Microscopy imaging

Background:

• High-throughput fluorescence microscopy is crucial for systems biology.
• Accurate cell localization and quantification are essential for analyzing cellular processes.
• Supervised deep learning excels at cellular instance segmentation but requires extensive annotated data.

Purpose of the Study:

• To develop a generalizable method for generating large instance segmentation training datasets for tissue-culture cells.
• To train vision transformer-based Mask R-CNN models for accurate cell segmentation and classification.
• To address class imbalance issues in biological datasets for improved model performance.

Main Methods:

• A novel method for generating synthetic instance segmentation training data for transmitted light microscopy images.
• Training of vision transformer (ViT) based Mask R-CNN models using generated datasets.
• Implementation of probabilistically weighted loss functions and partisan data collection to handle class imbalance between m-phase and interphase cells.

Main Results:

• Generation of large, high-quality instance segmentation training datasets for tissue-culture cells.
• Development of highly accurate object detectors capable of segmenting and classifying diverse cell types.
• Successful mitigation of dataset class imbalance, leading to improved model performance.

Conclusions:

• The proposed method provides a generalizable approach for creating large-scale training data for cell instance segmentation.
• The trained models demonstrate high accuracy in segmenting and classifying cells as m-phase or interphase.
• The methodology is adaptable to various adherent tissue culture cell lines, offering broad applicability in biological research.