Zebrafish neural imaging: event detection & source separation

Neural imaging gives you video of activity — and the hard part is deciding what belongs to the biology, what belongs to the measurement, and what evidence is safe to trust.

The problem: brief events in noisy video

Voltage and calcium imaging record brightness changes that correspond, imperfectly and on different time scales, to neural activity. The events you care about are localized and low-contrast, sitting on top of photobleaching, motion, persistent anatomy, and sensor noise. A naive brightness threshold either drowns in candidates or misses real activity.

So the real task is not 'find bright spots' — it is 'decide, per pixel and per frame, whether a change is more surprising than the background should produce.'

Current research: source separation without erasing the signal

The newest experiment track asks whether source-separation and denoising methods can improve event discovery in spontaneous calcium-burst recordings without changing the amplitude, footprint, or timing of the activity we ultimately want to interpret.

The clearest result so far is architectural rather than a single winning model: quiet-field standardization and spatial context have mattered more than model depth. At a budget of 58 candidates per burst, nested single-feature selection recovered 58 of 79 known events, compared with 51 of 79 for the archived carrier. At tighter budgets of 20 and 40, however, the native standardized carrier still performed best.

That trade-off matters. Several PCA, ICA, factorization, spectral, and state-space methods reduced noise or improved ranking, but none simultaneously passed the full preservation audit. The working design therefore keeps the activity carrier intact, uses separation outputs as bounded auxiliary evidence, and ranks candidates with a separate score.

Read the experiment note →

Preserve

Keep an immutable activity carrier for scientific traces, timing, and downstream interpretation.

Separate

Use spatial, spectral, ICA, factorization, and dynamical channels as bounded auxiliary features.

Rank

Compare candidate ordering at fixed budgets instead of treating a larger candidate list as better precision.

Gate

Advance only when detection improves without unacceptable amplitude, area, timing, morphology, or stability loss.

Borrowing from radar

The detectors here come from radar signal processing, where the same problem — spot a faint target against a fluctuating background — has decades of theory behind it. I use CFAR-style adaptive thresholding to hold the false-alarm rate stable as the background changes, and a Quadratic Gamma Discriminator to flag deviations from a learned noise model.

Treating a microscopy problem as a detection problem is the whole idea: it means the thresholds adapt instead of being hand-tuned per recording.

The detection pipeline

Detection is built as a set of interchangeable stages so the pre-processing filter and the detector can be swapped and tuned independently. A parallel grid search runs across filter and detector parameters, and the best operating points are chosen by real detection-theory metrics rather than by eye.

Pre-process (swap)

A Gamma spatio-temporal filter for enhancement, or Kalman–MCC for robust background estimation via maximum correntropy — chosen per recording.

Detect

A CFAR detector scores each pixel/region over time and surfaces candidate events.

Evaluate + select

FROC (true-positive rate vs. false positives per image) and truncated AUC to compare detectors across the search, then top-K model summaries.

Report

Per-model diagnostics: event-duration histograms, FROC curves, sensitivity heatmaps, PSD comparisons, and qualitative montages.

Keeping a human in the loop

Detectors are only half the system. A separate review workbench (a Fiji/Groovy plus browser workflow) lets a scientist inspect regions of interest and events with zoom, contrast, and overlay controls, accept/reject/mark-unsure at both the ROI and event level, and see a robust Kalman baseline under each trace. Decisions autosave, and ROI and event tables export for downstream analysis — so human judgment feeds back into which parameters win.

Modeling activity on a template grid

Beyond detection, a template-aligned 32×32 grid workflow models zebrafish left / right / neutral activity end to end: a grid autoencoder compresses each frame, a GRU predicts the latent code forward in time, and a classifier reads those latent codes — all with video-level splits so evaluation never leaks across a recording.