Iterative unsupervised domain adaptation for generalized cell detection from brightfield z-stacks.

Area of Science:

• Computational Biology
• Biomedical Imaging
• Machine Learning in Life Sciences

Background:

• Accurate cell counting is crucial for biological and biomedical research, particularly for analyzing cell growth using brightfield microscopy.
• Deep learning offers high accuracy in cell detection but traditionally requires extensive manually annotated training data for each cell line.
• Existing methods struggle to generalize effectively to unseen cell lines that differ significantly from the training data.

Purpose of the Study:

• To develop a generalized cell detection method that minimizes the need for manual annotations for new cell lines.
• To improve the accuracy and generalizability of deep learning models for brightfield-based cell counting across different cell lines.
• To enable accurate cell detection in unseen cell lines without requiring specific annotations for them.

Main Methods:

• Proposed an iterative unsupervised domain adaptation technique to enhance model generalization.
• Utilized a U-Net-based deep learning model trained on three consecutive focal planes from brightfield image z-stacks.
• Leveraged high-precision predictions on unseen cell lines to automatically generate new training data for model refinement.

Main Results:

• The model, initially trained on PC-3 cells, showed improved detection accuracy on unseen cell lines (LNCaP, BT-474, 22Rv1) after domain adaptation.
• Achieved a significant increase in F1-score from 0.65 (supervised training) to 0.84 after unsupervised domain adaptation for 22Rv1 cells.
• Demonstrated a mean accuracy of 0.87 across target domains, with an average improvement of 16 percent due to the proposed method.

Conclusions:

• The developed method enables accurate detection of diverse cell lines from brightfield images using a single trained model.
• New cell lines can be effectively incorporated into the model without manual annotation, thanks to iterative domain adaptation.
• The approach significantly enhances the model's readiness for high-accuracy cell detection in previously unseen cell types.