Towards Neuromorphic Event-Based Sensing for High-Speed Multi-Spectral Classification and Tracking of Microparticles

Technical Summary: Towards Neuromorphic Event-Based Sensing for High-Speed Multi-Spectral Classification and Tracking of Microparticles

Problem Statement
Conventional image-based microfluidic systems face a fundamental trade-off between throughput, imaging speed, and data bandwidth. High-speed frame-based imaging (CMOS/CCD) often results in motion blur, reduced signal quality, and prohibitive data volumes that hinder real-time processing, particularly when monitoring high-velocity flows. While existing solutions like stroboscopic illumination or deep learning-based reconstruction have improved image quality or throughput, they remain tied to global frame constraints or introduce significant computational overhead. Furthermore, previous event-based sensing (EVS) applications in microfluidics have largely been restricted to tracking monodisperse, spherical particles or rely on learning-based inference, failing to address the morphological and optical complexity of real-world, polydisperse analytes. Crucially, the integration of spectral information for opaque analytes within event-based microfluidic architectures remains largely unexplored.

Methodology
The authors developed a neuromorphic microfluidic sensing platform designed for the simultaneous tracking, sizing, and color classification of polydisperse, irregular microparticles. The system architecture integrates three core components:

1. Microfluidic and Optical Setup: A microfluidic channel (ibidi µ-Slide) is illuminated by white LEDs in a reflection configuration to detect intrinsic color and surface reflectance, avoiding the "shadow effect" common in transmission setups. A 10x objective collects backscattered light, which is split via a beam splitter to two sensors: a conventional CMOS camera for ground-truth reference and a high-resolution event-based camera (IDS uEye XCP-E).
2. Spatially Multiplexed Spectral Encoding: A custom filter assembly composed of three Lee photographic filters (Tokyo Blue, Primary Green, Marius Red) is placed directly in front of the event-based sensor. These filters cover non-overlapping regions of the sensor, allowing spectral identity and motion to be encoded directly at the sensing stage without image reconstruction.
3. Signal Processing Pipeline: The asynchronous event stream is temporally discretized into frames (at 500, 1000, and 1500 Hz) and spatially aggregated into 10×10 bins.
   • Denoising: A neighbor-count filtering strategy (8-connectivity kernel) removes isolated noise events, retaining only spatially clustered signals.
   • Color Prediction: A "Best-of-Frames" maximum-likelihood strategy is employed. For each spectral band (R, G, B), the algorithm identifies the peak event activity score across the entire transit duration, independent of the particle's exact position at any single moment. The final classification is the spectral region yielding the highest integrated peak.
   • Kinematic Profiling: Particle trajectories are reconstructed by linking centroids of filtered event clusters. Velocity is calculated from displacement between centroids, and size is estimated using an equivalent ellipse model derived from a weighted spatial covariance matrix, utilizing the minor axis for robustness against motion-induced distortion.

Key Contributions

• First Integration of EVS and Multi-Spectral Sensing in Flow: The work demonstrates, for the first time, the integration of event-based spectral sensing with continuously flowing, heterogeneous samples in a microfluidic channel.
• Label-Free, Training-Free Classification: The system extracts spectral signatures and motion directly from the sensing stage, eliminating the need for image reconstruction or complex, training-based neural network inference.
• Handling of Real-World Heterogeneity: Unlike prior EVS studies focused on idealized spherical beads, this platform successfully characterizes polydisperse, irregularly shaped microparticles (created via mechanical abrasion of thermoplastic polymers) under non-laminar flow conditions.
• Data Efficiency: The event-driven architecture drastically reduces data bandwidth by decoupling temporal resolution from data volume.

Results
The system was validated using red, green, and blue microplastic fragments (0.08–0.25 mm) flowing at velocities up to ~80 mm/s.

• Classification Performance: At an optimal frame construction frequency of 1000 Hz, the system achieved a global color classification accuracy of 74%. For the smaller particle size range (0.08–0.18 mm), accuracy reached 82%. Performance decreased for larger particles and at lower frame rates (500 Hz) due to temporal undersampling and spatial mixing of spectral regions.
• Kinematic Profiling: The system successfully resolved particle trajectories and velocities. Size estimation using the minor axis of fitted ellipses showed close agreement with CMOS ground truth, even under non-laminar flow and varying focus conditions.
• Throughput and Bandwidth: The system sustained an area throughput of 460 mm²/s. Compared to a conventional 1000 FPS CMOS camera, the event-based approach reduced data bandwidth by >240× (from ~3.84 GB/s to ~16 MB/s), primarily by eliminating redundant background data.
• Robustness: The platform maintained performance under non-ideal conditions, including air bubbles, particle rotation, and varying flow velocities, demonstrating resilience to the stochastic nature of real-world microfluidic flows.

Significance and Claims
The authors position this work as a proof-of-concept for a scalable, computationally efficient solution for high-speed, label-free screening of heterogeneous analytes. They emphasize that the results were achieved using off-the-shelf components (commercial RGB filters, standard white LEDs) and deliberately non-ideal conditions (non-laminar flow, irregular particles). Consequently, the reported performance is presented as a conservative baseline. The paper claims that this neuromorphic framework paves the way for future applications in clinical diagnostics (e.g., single-cell analysis) and environmental monitoring (e.g., microplastic detection) by overcoming the motion blur and data bottlenecks of traditional frame-based systems. The authors note that while the current setup uses broad-spectrum filters, future iterations could employ dedicated optical designs and tailored illumination to further improve classification accuracy and reduce spectral cross-talk.