Neuromorphic imaging flow cytometry combined with adaptive recurrent spiking neural networks

Area of Science:

• Neuromorphic Engineering
• Biomedical Imaging
• Computational Neuroscience

Background:

• Traditional flow cytometry faces limitations in real-time, high-resolution analysis.
• Spiking neural networks (SNNs) offer potential for low-latency, sparse data processing.

Purpose of the Study:

• To develop and evaluate an imaging flow cytometer with neuromorphic processing for enhanced particle classification.
• To assess the performance of recurrent and feedforward SNNs in classifying microparticles.

Main Methods:

• Utilized a 1 µs temporal resolution event-based CMOS camera for high-speed imaging.
• Employed adaptive feedforward and recurrent spiking neural networks (SNNs) for data analysis.
• Classified polymethyl methacrylate (PMMA) particles of varying sizes (12-20 µm) in a microfluidic channel at 0.7 m/s.

Main Results:

• Recurrent SNNs (LSTM, GRU) achieved 98.4% classification accuracy by exploiting temporal dependencies.
• Adaptation mechanisms in feedforward SNNs improved accuracy by 4.3%.
• The system demonstrated low latency and sparse data processing capabilities.

Conclusions:

• Neuromorphic-assisted imaging flow cytometry significantly enhances classification performance.
• SNNs, particularly recurrent architectures, are effective for high-speed particle analysis.
• This technology provides a roadmap for advanced neuromorphic applications in biomedicine.