Event-Based Imaging Cytometry Combined with Recurrent/Feedforward Adaptive Spiking Neural Networks

Imaging flow cytometry (IFC) is a cellular analysis technique that combines the features of flow cytometry and confocal microscopy to provide rich morphological information. However, typical IFC schemes are hindered by limited spatial resolution caused by motion blur as well as the substantial data volume required, imposing restrictions in storage and processing. Event cameras address these limitations by asynchronously measuring brightness changes per pixel, offering high temporal resolution and spiking sparse data. This IFC modality has been combined with various deep neural network as processing back-end, but despite their efficiency, conventional networks cannot exploit the asynchronous spiking nature. In this work, we utilize experimental data from an event-based IFC system [1] targeting the classification of three particle classes, having different diameters of 12 μm, 16 μm and 20 μm, The proposed back-end in our case consists of lightweight feedforward and recurrent spiking neural networks, aiming to reduce complexity, latency and exploit the inherent spiking input. In addition, adaptive spiking mechanisms are incorporated to our leaky integrate and fire (LIF) neurons boosting classification accuracy by >3% while marginally increasing the number of multiplications during inference.