aiSEGcell: User-friendly deep learning-based segmentation of nuclei in transmitted light images

Area of Science:

• Cellular imaging
• Bioimage analysis
• Machine learning in biology

Background:

• Cellular structure segmentation in microscopy typically requires fluorescent labeling.
• Fluorescent labeling can increase phototoxicity, limit imaging channels, and slow down acquisition.
• Current automated methods often need specialized expertise and optimized conditions.

Purpose of the Study:

• To introduce aiSEGcell, a user-friendly software for segmenting cells and nuclei in bright field microscopy images using CNNs.
• To evaluate aiSEGcell's performance across diverse cell types and imaging modalities.
• To provide a large, curated dataset for training and validation.

Main Methods:

• Development of a CNN-based software, aiSEGcell, for bright field image segmentation.
• Extensive evaluation on 2D cell cultures from various imaging modalities.
• Creation and provision of a ground-truth dataset comprising 1.1 million nuclei in 20,000 images.
• Retraining capability demonstrated with as few as 32 images.

Main Results:

• aiSEGcell accurately segments nuclei in challenging bright field images, comparable to manual segmentation.
• The software preserves biologically relevant information for quantitative analysis, including noisy biosensor data.
• Demonstrated high adaptability for new segmentation tasks with minimal retraining data.

Conclusions:

• aiSEGcell offers an accessible and efficient solution for cell and nucleus segmentation in bright field microscopy.
• The software reduces reliance on fluorescent labeling, thereby minimizing phototoxicity and speeding up imaging workflows.
• aiSEGcell empowers researchers without coding expertise to perform advanced bioimage analysis via command-line or GUI.