Machine learning of diffraction image patterns for accurate classification of cells modeled with different nuclear...

Area of Science:

• Biomedical Optics
• Computational Biology
• Cellular Imaging

Background:

• Accurate measurement of nuclear-to-cytoplasm (N:C) ratios is crucial for identifying atypical and tumor cells.
• Current clinical methods often depend on immunofluorescence staining and manual analysis, which can be time-consuming and subjective.

Purpose of the Study:

• To develop a rapid, label-free method for cell classification using optical cell models (OCMs) and diffraction imaging.
• To evaluate the effectiveness of simulated diffraction patterns in distinguishing cells based on N:C ratios.

Main Methods:

• Generated 1892 realistic OCMs with varying nuclear volumes and orientations.
• Simulated cross-polarized diffraction image (p-DI) pairs for OCMs categorized into small (OCM_S), medium (OCM_O), and large (OCM_L) nuclear size groups.
• Extracted image parameters using the gray-level co-occurrence matrix algorithm for binary classification with a support vector machine (SVM) classifier.

Main Results:

• Achieved high average accuracies for SVM binary classification: 98.8% (OCM_S vs OCM_O) and 97.5% (OCM_O vs OCM_L) for prostate cancer cell structures.
• Similar high accuracies were observed for smaller prostate normal cell structures (98.9% and 97.8%).
• Demonstrated robust performance of SVM compared to clustering classifiers, indicating the utility of diffraction pattern correlations.

Conclusions:

• High-order correlations in diffraction patterns are effective for label-free detection of single cells with large N:C ratios.
• The developed method shows potential as a rapid and accurate alternative to traditional cell classification techniques.
• This approach could significantly advance early disease detection and cellular analysis.