Cell-MICS: Detecting Immune Cells With Label-Free Two-Photon Autofluorescence and Deep Learning.

Area of Science:

• Biomedical Optics
• Computational Biology
• Immunology

Background:

• Multiphoton imaging offers deep-tissue visualization with metabolic contrast, valuable for studying inflammation.
• Label-free two-photon autofluorescence (2P-AF) lacks specificity compared to antibody-based methods.
• Identifying specific immune cells in unstained tissues remains a challenge for current imaging techniques.

Purpose of the Study:

• To investigate the potential of multiphoton imaging with computational specificity (MICS) for reliable immune cell classification.
• To develop and validate a deep learning model for distinguishing immune cell types using label-free 2P-AF.
• To assess the contribution of metabolic cofactors (NADH and FAD) to classification accuracy.

Main Methods:

• Training a convolutional neural network (CNN) utilizing a low-complexity SqueezeNet architecture on images of immune cells.
• Evaluating classification performance using metrics such as ROC-AUC, PR-AUC, F1 score, precision, and recall.
• Conducting perturbation tests to assess model robustness against extracellular environmental factors.

Main Results:

• The CNN achieved high accuracy in binary classification (0.89 ROC-AUC, 0.95 PR-AUC) between T cells and neutrophils.
• The model demonstrated moderate performance in multi-class classification of six isolated cell types (0.689 F1 score).
• Perturbation tests confirmed the model's independence from the extracellular environment, highlighting the equal importance of NADH and FAD autofluorescence.

Conclusions:

• Deep learning can provide computational specificity for identifying immune cells in label-free multiphoton imaging.
• MICS holds significant potential for advancing unstained tissue analysis and in vivo endomicroscopy.
• This approach could revolutionize label-free inflammation research by enabling precise immune cell identification.