Machine learning issues and opportunities in ultrafast particle classification for label-free microflow cytometry

Area of Science:

• Biophysics
• Optical Engineering
• Machine Learning

Background:

• Label-free imaging microflow cytometry enables high-throughput single-particle analysis.
• Current machine learning methods face limitations in computational cost, data storage, and susceptibility to training bias.

Purpose of the Study:

• To introduce a computationally efficient and versatile machine learning approach for microparticle classification.
• To address and mitigate bias in machine learning models used in flow cytometry.
• To demonstrate proof-of-principle classification using a novel label-free microflow cytometer.

Main Methods:

• Development of a simple, low-cost, label-free microflow cytometer.
• Implementation of a machine learning algorithm for analyzing interference patterns of flowing microparticles.
• Investigation of bias detection and prevention strategies for machine learning models under varying conditions.
• Exploration of diffraction gratings to modify particle patterns for optical extreme learning machines.

Main Results:

• Successful proof-of-principle classification of PMMA microbeads ([Formula: see text] and [Formula: see text] diameters).
• Demonstration of extremely low computational cost for microparticle classification.
• Exhibited good generalization across variations in particle position.
• Detailed discussion on detecting and preventing machine learning bias.

Conclusions:

• The proposed machine learning approach offers a viable solution for high-throughput, low-cost microparticle analysis in label-free imaging flow cytometry.
• The method shows robustness against variations in measurement conditions and particle position.
• Further research into optical extreme learning machines with diffraction gratings is warranted.