Scan-Direction Bias: How Event-Camera Smear Corrupts Live Atlas Registration in Fast Volumetric Imaging

Event cameras record only the pixels whose brightness changes, and each such report is called an event. This makes them attractive for whole-organ microscopy in which the specimen moves continuously under the objective: every tissue edge sweeps across the sensor and fires events, so the camera's usual blindness to static scenes disappears. The same physics also creates a hazard. In dim light, each pixel's amplifier responds more slowly, so events trail behind moving edges. This smear points along the scan direction, and its length depends on local brightness. When such data is mapped onto a standard brain atlas, the smear does not look like noise. It looks like a real shift or shear of the anatomy, and the registration software absorbs it into the transform. We propose to measure this bias, to correct it with a calibration that comes free from ordinary back-and-forth scanning, and to quantify what it costs in the setting where it matters most: registration performed live, during acquisition, to decide where to spend slow high-resolution imaging time.

The second half of the idea borrows a lens from clinical AI. Mind the Performance Gap showed that a hospital model lost accuracy after deployment mainly because the live data pipeline differed from the curated research warehouse ("infrastructure shift", −0.016 AUROC), not because patients changed over time (+0.005 AUROC). Fast volumetric imaging has the same split. Methods are benchmarked on curated, fully stitched volumes that were registered offline, but an adaptive microscope must register strip by strip while the scan is still running. We apply the same three-snapshot decomposition to imaging sessions. The goal is to find out whether the live pipeline, rather than the sensor, is the dominant source of misregistration, and whether correcting the smear closes that gap.

Core Claim

In continuous-scan event-based imaging, amplifier latency produces an edge displacement that is approximately

$$\delta(I) = v\,\tau(I),$$

where $v$ is stage velocity and $\tau(I)$ is the pixel response latency, which grows as local intensity $I$ falls. The latency model comes from v2e, which simulates intensity-dependent pixel bandwidth and shows that the "event cameras have no motion blur" assumption fails in low light. From Sim-to-Real makes the same point.

Our claim has three parts. First, $\delta(I)$ is large enough, at the short dwell times high-throughput imaging needs, to bias atlas registration by a measurable amount. Second, the bias flips sign when the scan direction reverses. Third, it can be estimated without extra imaging time, because a serpentine (back-and-forth) scan places adjacent strips in opposite directions, and their overlapping seam shows an offset of $2v\tau(I)$. Fitting $\tau(I)$ to the seam offsets binned by local event rate yields a per-session latency model. That model lets us shift events back to their true positions before reconstruction and registration.

What is new here is the following. Prior event-denoising work treats low-light corruption as random background activity, to be deleted. E-MLB benchmarks eleven such denoisers, and S2N characterizes the dual degradation of signal loss plus noise. We instead treat the smear as a deterministic, direction-signed geometric error, and we evaluate its consequence in atlas coordinates rather than event-level denoising scores.

Relation to Prior Work

Two papers motivate the setting by showing that event sensors earn their place in volumetric microscopy. EventLFM pairs an event camera with Fourier light-field optics for 3D imaging up to 1 kHz, with better signal-to-background ratio in scattering mouse brain tissue. Inter-event Interval Microscopy recovers static fluorescence by pulsing the illumination. It also shows that the interval between events encodes brightness, a fact we reuse as the intensity proxy for $\tau(I)$. Both papers reconstruct in raw sensor coordinates, never register to an atlas, and name event noise as an open problem. Our proposal takes their sensor and adds the coordinate step they skip. We replace their light-field and modulation tricks with continuous stage motion of the kind VISOR uses to image a mouse brain in under 1.5 hours.

Two papers supply the case that atlas alignment is worth getting right:

Both assume clean frame-camera inputs. Neither asks what happens when the input carries a structured geometric bias.

The application regime comes from Serial optical coherence microscopy. There, a whole mouse brain takes under an hour at standard resolution but roughly ten hours at high resolution. A fast first pass followed by targeted high-resolution re-imaging of chosen atlas regions is therefore valuable, but only if the live registration that chooses the regions is correct.

The evaluation methodology comes from Ötles et al.'s decomposition. The broader prospective-validation literature adds a constraint: what matters is the downstream decision, not standalone accuracy. That literature includes AI support for COVID-19 CT diagnosis and prospective AI-assisted breast screening. For that reason, our headline metric is whether the high-resolution pass lands in the intended region. We do not run a clinical trial; the pathology connection is the motivating use, not an evaluated claim.

Research Program

Rig. We build a continuous-scan rig with two cameras viewing the same field through a beam splitter. One is a Prophesee EVK4 event camera, the same sensor as the IEIM work. The other is a co-aligned sCMOS camera that provides a ground-truth intensity image of every strip. Specimens are fixed fluorescently labelled mouse brain sections and thick slabs. We vary stage velocity and illumination power, and add neutral-density attenuation at the E-MLB transmittance levels (1/4, 1/16, 1/64). This lets us sweep the flux-times-velocity regime in which smear should appear.

Simulation arm. In parallel, v2e converts the sCMOS strips into synthetic events under controlled latency settings. This lets us check whether the seam-calibrated $\tau(I)$ recovers the latency that was simulated.

Three components. The first is the measurement of bias. We register event reconstructions and sCMOS reconstructions of the same strips to the Allen Common Coordinate Framework (the standard mouse atlas cited by the International Brain Laboratory). We then report the difference in recovered transforms as a function of direction, velocity and flux.

The second is the seam-calibrated correction. We fit $\tau(I)$ from serpentine overlaps, using inter-event intervals as the local intensity estimate. We then displace events by $-v\tau(I)$ along the scan axis before reconstruction.

The third is the live-versus-offline decomposition. It uses three snapshots:

  • $D_{ret}$ is archived sessions processed offline with full-volume denoising and global registration.
  • $D_{pro}$ is the output the online pipeline produced during new sessions. This pipeline does strip-wise streaming registration with only past strips available, and it is subject to real dropped packets and bandwidth saturation.
  • $D'_{ret}$ is those same new sessions reprocessed offline afterward.

The gap between $D_{pro}$ and $D'_{ret}$ isolates the pipeline (infrastructure) component. The gap between $D'_{ret}$ and $D_{ret}$ isolates acquisition drift, such as tissue batch and sensor temperature.

Baselines

The naive reference point accumulates raw events into frames per strip and stitches them with standard intensity-based affine-plus-B-spline registration to the atlas. It applies no denoising and no latency handling, and it is run identically in online and offline modes.

The first denoise-then-register baselines use the two classical filters E-MLB benchmarks, background activity filtering and KNoise. The second uses EDnCNN, one of the learned denoisers E-MLB reports as best on its Event Structural Ratio score; the summary in our file gives no numeric value for that score. All three share the same reconstruction and registration code as our method, so only the event preprocessing differs.

The strongest event-side competitor is ProgressiveMotionSeg. It jointly estimates motion by maximizing image sharpness and denoises events over seven iterations. Because sharpness maximization could in principle absorb the smear, it is the method most likely to make our correction unnecessary. On ESIM synthetic data at noise level 0.25 it reports 79.52% mean intersection-over-union, against 60.61% for EBMS. We supply it with the known stage velocity, and it receives the same compute budget per strip as our pipeline in the live setting.

A learned alternative replaces the classical registration with a feature-space alignment module of the kind used by Li et al. It is trained on rig data with the same tuning budget as the other learned components.

The strongest non-event competitor is the co-aligned sCMOS reconstruction run through the identical registration pipeline. This is the VISOR-style answer to whether event cameras are needed at all.

The decisive ablation is directional. We give the naive pipeline a uniform translation correction fitted per session. This tests whether a flux-independent offset is enough, or whether the intensity dependence of $\tau(I)$ is the ingredient that matters.

No reported result was found on Lacuna for atlas registration error of any event-camera reconstruction, so all baselines must be run fresh.

"Beating" a baseline has three concrete thresholds:

  • Our corrected pipeline achieves lower target registration error than every event baseline, measured at manually annotated landmarks against the sCMOS-derived registration.
  • It reaches within the sCMOS reference's own landmark annotation variability.
  • It raises the re-imaging hit rate, meaning the fraction of the high-resolution pass falling inside the intended atlas region, by more than the run-to-run spread across repeated sessions.

Why not just…?

Why not just use an sCMOS or line-scan camera, as VISOR does? This is a serious alternative, and the experiment answers it directly, because the sCMOS arm sees the same specimens through the same optics. The case for event sensors rests on two documented advantages: higher signal-to-background ratio in scattering tissue in EventLFM, and wider dynamic range than a saturating sCMOS in IEIM. Neither has been shown in a whole-organ continuous scan. If corrected event data registers no better than sCMOS at equal throughput, we will report that event sensors only pay off in thick, scattering slabs, where the background-ratio advantage should be largest. That would still be a useful, bounded result.

Why not just denoise harder before registering? Denoisers remove isolated background events. The smear, however, is dense and spatially correlated with real edges, which is exactly the kind of structure denoisers are built to keep. We therefore expect background filtering, KNoise and EDnCNN to leave the direction-signed bias intact. The decisive check is whether their registration error still flips sign with scan direction. If it does not flip, denoising suffices and our correction is unnecessary.

Why not just buffer the whole volume and register offline? For pure archiving this is correct, and the $D'_{ret}$ snapshot measures exactly how much it helps. It fails for adaptive acquisition, because high-resolution re-imaging regions must be chosen before the fast pass ends. The SOCM trade-off, under one hour versus roughly ten hours, is what makes those online choices valuable. If the $D_{pro}$-versus-$D'_{ret}$ gap turns out negligible, the live-pipeline story collapses to the bias story alone, and we will say so.

Why not just scan every strip in both directions and average? This cancels the first-order bias. It also costs twice the acquisition time, which defeats the purpose of high throughput. The seam calibration recovers the same information from the overlaps a serpentine scan already contains. We include double-coverage averaging as an oracle, and we expect our single-coverage correction to come within its error at half the time.

Decisive Checks

The idea is supported only if three predictions hold.

  • Sign flip. Uncorrected registration offsets between event and sCMOS reconstructions flip sign with scan direction.
  • Brightness scaling. Their magnitude grows with velocity and with dimmer illumination in the way v2e's latency model predicts. In simulation, the seam-fitted $\tau(I)$ recovers the latency that was injected.
  • Brightness dependence matters. After seam calibration, the residual error falls further than with the uniform-offset ablation.

The idea is falsified on either of two outcomes. One is that offsets show no directional sign and no brightness dependence at realistic dwell times. The other is that ProgressiveMotionSeg with known velocity already removes them.

The gap decomposition then tells us where future effort should go. If the pipeline component, rather than acquisition drift, dominates the live-versus-offline difference, as it did for Ötles et al.'s hospital data, the main engineering problem in fast volumetric imaging lies in the online pipeline more than in the sensor.