Different cell imaging methods did not significantly improve immune cell image classification performance.

Area of Science:

• Cell Biology
• Computational Biology
• Microscopy

Background:

• High-throughput microscopy generates vast cell image datasets requiring automated analysis.
• Machine learning, particularly deep learning, is crucial for extracting information like cell counting and classification.
• Systematic evaluation of imaging method effects on machine learning accuracy is lacking.

Purpose of the Study:

• To investigate the impact of various imaging methods on machine learning-based cell type classification accuracy.
• To compare the performance of convolutional neural networks (CNNs) across different microscopy techniques.

Main Methods:

• Observed lymphoid-primed multipotential progenitor (LMPP) and pro-B cells using differential interference contrast (DIC), phase contrast (Ph), and bright-field (BF) microscopy.
• Trained and evaluated CNNs for cell type classification using individual and combined imaging methods.
• Compared CNN performance against classifiers using only cell size or shape features.

Main Results:

• CNNs achieved high classification performance (Area Under the ROC Curve ~0.9), outperforming methods relying solely on cell size or shape.
• No significant differences in classification accuracy were observed between DIC, Ph, and BF imaging methods.
• Focal position did not significantly affect the performance of the machine learning classifiers.

Conclusions:

• Machine learning classifiers, specifically CNNs, demonstrate robust performance in cell type identification from microscopy images.
• The choice of imaging method (DIC, Ph, BF) and focal position has minimal impact on the accuracy of these automated cell classifiers.
• Automated cell image analysis using deep learning offers a reliable approach for large-scale biological data interpretation.