Deep learning of diffraction image patterns for accurate classification of five cell types.

Area of Science:

• Biophotonics
• Computational Biology
• Cell Biology

Background:

• Label-free cell classification is crucial for high-throughput biological research and clinical diagnostics.
• Existing methods often require cell labeling or lack sufficient accuracy and throughput.
• Advanced computational approaches are needed to extract meaningful features from complex cellular imaging data.

Purpose of the Study:

• To develop and validate a deep neural network (DINet) for label-free, high-throughput cell classification using polarized diffraction imaging.
• To assess the accuracy and robustness of DINet in classifying diverse cell types.
• To investigate the feature extraction mechanisms within DINet that contribute to accurate cell classification.

Main Methods:

• Development of DINet, a deep neural network designed to extract features from cross-polarized diffraction image (p-DI) pairs at multiple pixel scales.
• Measurement of 6185 cells using polarization diffraction imaging flow cytometry (p-DIFC).
• Classification of cells using DINet on p-DI data, with performance evaluated using 5-fold cross-validation.

Main Results:

• Achieved an average classification accuracy of 98.9% ± 1.00% on test datasets.
• Demonstrated invariance of DINet to image translation, rotation, and blurring.
• Identified that high-order correlations extracted in deeper convolutional layers (3 and 4) are critical for distinguishing between correctly and incorrectly classified cells.

Conclusions:

• DINet provides a highly accurate and robust label-free method for cell classification.
• The deep feature extraction capabilities of DINet are essential for accurate cell type identification.
• This approach holds significant potential for advancing cell analysis in both research and clinical settings.