Deep Cytometry: Deep learning with Real-time Inference in Cell Sorting and Flow Cytometry.

Area of Science:

• Biomedical engineering
• Machine learning
• Optical measurement technology

Background:

• Photonic time-stretch instruments generate high-throughput data (1 Tbit/s), ideal for deep learning applications.
• Previous methods combined time-stretch microscopy with feature extraction and deep learning for cell classification.
• These methods, while effective, involved computationally intensive signal processing and feature extraction.

Purpose of the Study:

• To develop a more computationally efficient deep learning pipeline for real-time cell classification.
• To eliminate the need for traditional signal processing and feature extraction steps.
• To enable low-latency cell sorting applications using deep learning.

Main Methods:

• A convolutional neural network (CNN) was designed to directly process raw measurement signals from time-stretch instruments.
• The CNN pipeline bypasses conventional image processing and feature extraction.
• The method was tested for label-free classification of OT-II white blood cells and SW-480 epithelial cancer cells.

Main Results:

• The new deep learning pipeline achieved cell classification in milliseconds.
• The CNN directly operated on measured signals, significantly improving computational efficiency.
• Classification accuracy for both cell types exceeded 95% in a label-free manner.

Conclusions:

• The developed deep learning pipeline offers a computationally efficient approach for high-throughput cell analysis.
• This method enables real-time cell sorting by providing rapid classification decisions.
• The technology shows significant promise for early cancer detection and advanced biomedical instrumentation.