High-Throughput Particle Identification in Acoustic Flow Cytometry Using RF Time-Series Learning
(Code and Paper will be released soon)
High-Throughput Particle Identification in Acoustic Flow Cytometry Using RF Time-Series Learning
(Code and Paper will be released soon)
Most methods for identifying microplastics rely on optical or spectroscopic analysis in a lab, which means capturing particles, preparing samples, and imaging them under controlled conditions. This study takes a faster route: identifying particles while they flow, directly from high-frequency ultrasound signals.
In acoustic flow cytometry, sheath flows focus particles into a single stream that passes through the focus of an ultrasound transducer. Each particle produces a short radio-frequency (RF) echo. Over three experimental runs, 5,100 of these signals were recorded from polystyrene (PS), polymethyl methacrylate (PMMA), and glass particles.
A two-stage pipeline then analyzes each signal:
Event filtering: Handcrafted signal features separate real particle events from background and air-bubble signals. Gradient boosting does this with 99.26% accuracy.
Classification: Valid particle signals are classified by material. A compact Transformer reaches 95.73% accuracy, matching the much heavier HIVE-COTE2 benchmark while cutting processing time per particle from about 56 seconds to 1.8 milliseconds.
Which signal features separate particles from noise?
SHAP (SHapley Additive exPlanations) shows how much each signal feature pushes the gradient boosting model toward each signal type.
Background: Signal energy arrives later and the waveform is more sharply peaked, so a high temporal centroid and high kurtosis are the strongest indicators.
Air bubble: Energy arrives early and the event is short, so a low temporal centroid and a short event duration point to a bubble.
Particle: A low temporal centroid and a low dominant frequency generally indicate a particle, but no single feature dominates. Particle signals depend on a combination of features, which is why the second stage classifies them directly from the raw RF waveform using deep learning.
Which parts of the signal reveal particle type?
Integrated Gradients shows which parts of an RF signal most influence the Transformer’s decision. For each particle type, the highlighted regions line up with the main particle echo, the strongest part of the signal.
PS: Attribution centres on the main echo and spreads across several neighbouring oscillations, not just the single highest peak.
PMMA: The strongest attribution matches the main echo, which arrives later in the signal, with smaller contributions from the waveform before it.
Glass: Attribution is more spread out but still peaks at the main echo.
The model uses the shape of the echo rather than a fixed time position, since the echo arrives at a different time for each particle type. Echo timing and strength also depend on how close the particle passes to the transducer, so these maps show what the model is sensitive to.