Image-based phenotyping of disaggregated cells using deep learning.

Area of Science:

• Cell biology
• Biotechnology
• Machine learning in medicine

Background:

• Cell phenotyping is crucial for biological research and medical diagnostics.
• Current fluorescence-based methods have limitations and are not always applicable.
• Existing machine learning approaches for image cytometry struggle with cell clumping and distinguishing subtle cell differences.

Purpose of the Study:

• To develop a reliable method for accurate cell phenotyping using low-resolution bright-field and non-specific fluorescence microscopy images.
• To assess the efficacy of machine learning in phenotyping disaggregated single cells, even those difficult to distinguish visually.

Main Methods:

• A convolutional neural network was trained using automatically segmented images of cells from eight standard cancer cell lines.
• The model utilized low-resolution bright-field and non-specific fluorescence images capturing the nucleus, cytoplasm, and cytoskeleton.
• The model's performance was validated on separately acquired image datasets.

Main Results:

• The developed machine learning model achieved a high average F1-score of 95.3% in phenotyping cancer cell lines.
• Accurate cell identification was demonstrated even with low-resolution and non-specific imaging.
• The approach successfully phenotyped disaggregated single cells, overcoming previous limitations.

Conclusions:

• Machine learning applied to basic microscopy images enables accurate cell phenotyping.
• This technique offers a promising alternative to complex fluorescence-based methods.
• The development of an "electronic eye" for direct microscopy image-based cell phenotyping is feasible.