Drugs that stop inflammatory organ damage work by blocking leukocyte recruitment — the cascade in which circulating white blood cells slow down, roll along the vessel wall, stick firmly, and crawl out into tissue. But almost everything we know about these drugs comes from snapshots taken after the fact: gene-expression counts of neutrophil markers, post-mortem cell counts, blots for the fibers neutrophils spit out. The live movie — which step of the cascade a drug actually interrupts, in which vessel, at which hour after injury — is missing, because the organs where recruitment matters most (a beating heart in sepsis, a breathing lung, a pulsating brain cortex) move far more than the cells being watched. In intravital microscopy, millimeter-scale tissue excursions from heartbeat and respiration swamp micrometer-scale leukocyte displacements, and the motion's frequency overlaps the signal's, so filters cannot separate them.
The signal-processing world has already solved an isomorphic problem in wearables and radar: recovering a sub-millimeter pulse from a walking person by learning to discard motion-correlated features rather than by filtering frequencies. I propose to port that disentanglement principle into intravital microscopy, as a phase-conditioned non-rigid deformation model fused with multi-object cell tracking, and to use it as a kinetic assay for leukocyte recruitment inhibition. The payoff is twofold: a measurement instrument the biology currently lacks, and a way to answer an open therapeutic question the static assays cannot — whether recruitment blockers still work when given after the insult rather than before.
Core Claim
Tissue motion in vivo is quasi-periodic and largely predictable from the combined cardiac and respiratory phase; leukocyte motion is aperiodic and goal-directed. This difference is exploitable. An image sequence of a vascular bed can be modeled as a static tissue template warped by a phase-conditioned deformation field, plus a small number of moving cells, plus noise:
$$I(x,t) \;=\; T\big(\Phi(x;\,\varphi(t))\big) \;+\; \sum_i C_i\big(x - p_i(t)\big) \;+\; \varepsilon,$$
where $\varphi(t)$ is the physiological phase (a two-dimensional cycle of cardiac and respiratory phase), $\Phi$ is a non-rigid deformation field conditioned on that phase, $T$ is the undeformed tissue, and $C_i$ are cell appearances at positions $p_i(t)$. The model is trained jointly for registration and tracking, with an adversarial penalty in the spirit of the motion discriminator used for radio respiration monitoring by Choi et al. in Remote Respiration Monitoring of Moving Person Using Radio Signals: a small network tries to predict $\varphi(t)$ from the residual cell-track features, and the tracker is penalized whenever it succeeds, forcing cell motion to be statistically independent of tissue motion. The claim is that this yields per-cell trajectories (rolling velocity, firm-adhesion dwell times, crawling, diapedesis events) in vascular beds that today cannot be imaged continuously, and that those trajectories reveal where in the recruitment cascade an inhibitor acts.
This is different from the closest existing work on every axis that matters. Fu et al. in Four Dimensional Image Registration for Intravital Microscopy built a registration pipeline for exactly this kind of data — intra-volume B-spline correction plus inter-volume alignment — but their temporal step is strictly rigid, they explicitly cannot handle continuous non-rigid organ deformation, they track nothing, and they touch no pharmacology. The cell-tracking benchmarks, Maška et al.'s The Cell Tracking Challenge: 10 years of objective benchmarking and Anjum & Gurari's CTMC: Cell Tracking with Mitosis Detection Dataset Challenge, evaluate trackers on sequences with no tissue motion at all. And the recruitment-inhibition studies — Li et al.'s Ferrostatin-1 in Ferrostatin-1 improves acute sepsis-induced cardiomyopathy via inhibiting neutrophil infiltration through impaired chemokine axis and Kim et al.'s exercise preconditioning in Exercise preconditioning alleviates photothrombotic ischemic stroke in mice by orchestrating neutrophils — are precisely the studies whose conclusions would change if the intermediate kinetics were visible; Li et al. show the chemokine network shifts within six hours but cannot say whether neutrophils then fail to roll, roll without arresting, or arrest and detach.
Research Program
The program has three parts, ordered so that each de-risks the next.
The first part builds and validates the model entirely on existing public data, before any animal work. The architecture is a 3D encoder-decoder that predicts the deformation field $\Phi$ conditioned on an estimated phase $\varphi(t)$ (estimated either from the image sequence itself via a periodicity head, or, in the animal, from external triggers), followed by a detection-driven multi-object tracker operating on the motion-corrected frames. Two design choices are borrowed with credit: the adversarial motion-discarding objective from Remote Respiration Monitoring of Moving Person Using Radio Signals, and the idea that a residual signal sharing coherent spatiotemporal structure across regions is the physiological one, from Li's Contrast-Phys in Contrast-Phys: Unsupervised Video-Based Remote Physiological Measurement via Spatiotemporal Contrast — here the coherent structure is the periodic tissue warp, which should be shared across the field of view while individual leukocyte tracks are not. For the encoder, I would initialize from published single-cell microscopy representations such as CaicedoLab/CHAMMI-75 rather than ImageNet weights, since intravital fluorescence contrast statistics are closer to that domain; low-contrast frames can be pre-denoised with a zero-shot denoiser in the style of MeijiTian/ZS-Denoiser-HREM, noting that Fu et al. had to hand-tune Gaussian smoothing and histogram equalization precisely because raw live-tissue frames are low-contrast and noisy.
The second part applies the model to the one existing dataset at this intersection: the intravital mouse-kidney time-lapse of immune cells across 61 time points and 4 spectral channels from Four Dimensional Image Registration for Intravital Microscopy. This tests whether phase-conditioned non-rigid correction beats their rigid inter-volume alignment at preserving track continuity, using their own reported improvement in per-pixel sum-of-squared differences — between 7.8% and nearly 30% across time points — as the registration-quality reference point.
The third part is the biological payload. Using the cecal ligation and puncture sepsis model and Ferrostatin-1 dosing from Ferrostatin-1 improves acute sepsis-induced cardiomyopathy via inhibiting neutrophil infiltration through impaired chemokine axis, image the coronary microvasculature through an epicardial window and quantify neutrophil kinetics in untreated, pre-treated, and post-onset-treated animals (dosing at the six-hour mark where Li et al. locate the chemokine effect, and later, to probe the open problem of the post-insult therapeutic window). The pre-registered hypotheses are mechanistically sharp: if Fer-1 acts by blinding chemokine gradients, it should preferentially reduce firm arrest and crawling while leaving rolling velocity comparatively intact; a rolling defect would instead implicate the selectin/adhesion axis described in Neutrophil, Not Macrophage, Infiltration Precedes Neointimal Thickening in Balloon-Injured Arteries. A secondary cohort applies the same assay to the photothrombotic-stroke model of Exercise preconditioning alleviates photothrombotic ischemic stroke in mice by orchestrating neutrophils, where a pial window exposes the cortex whose post-stroke neutrophil burden Kim et al. measured only by terminal Ly6G qPCR and immunofluorescence.
Baselines
The naive reference point is doing nothing: run a standard multi-object tracker directly on uncorrected frames. This is not a strawman — on genuinely hard, low-contrast microscopy video, the CTMC benchmark reports that Tracktor with a Faster-RCNN detector achieves a MOTA of only 23.59 CTMC: Cell Tracking with Mitosis Detection Dataset Challenge, which is the realistic floor for tracking through uncorrected physiological motion.
The second baseline is detect-and-discard: a frame-level artifact classifier in the spirit of artiNet, which reached an AUC of 0.90 on expert-annotated confocal endomicroscopy artifacts versus 0.85 for a correlation-angle feature with random forest and roughly 0.73 for HOG Motion Artifact Detection in Confocal Laser Endomicroscopy Images, and of Tiny-PPG, which achieved an 87.4% DICE coefficient for sample-wise artifact segmentation on PPG-DaLiA Tiny-PPG: A lightweight deep neural network for real-time detection of motion artifacts in photoplethysmogram signals on edge devices. Corrupted frames are flagged and removed, and tracking runs on the survivors. This baseline matters because it is the cheap thing a biologist would actually do, and because it fails in a measurable way: dropped frames break trajectory continuity, and on CTMC, tracking identity collapses when inter-frame association is disrupted.
The strongest existing-method baseline is the two-stage pipeline of Fu et al. Four Dimensional Image Registration for Intravital Microscopy — their axial interpolation, intra-volume B-spline correction, and rigid 4D inter-volume alignment — used as preprocessing ahead of the same tracker and the same detector backbone that my model uses. It is the strongest competitor because it is the only published method validated on intravital data with physiological motion; no reported tracking result exists on Lacuna for it, so the comparison will be run head-to-head on identical frames. Fairness is enforced by fixing the detector and tracker architectures, the training data, and the tuning budget across all baselines and the proposed model, and by evaluating on identical held-out sequences. Beating the baselines means three things jointly: higher track-continuity scores (IDF1 and identity switches) than rigid-preprocessing-plus-tracker at matched detection quality; registration quality at least matching the 7.8–30% SSD improvement Fu et al. report; and, most distinctively, passing a phase-invariance test — after correction, firm-adhesion dwell times and arrest counts must not correlate with the cardiac/respiratory phase at which they were measured. The oracle upper bound is DeepSORT run with ground-truth detections, which on CTMC reaches a MOTA of 99.32 and an IDF1 of 94.23 CTMC; no real system will reach that under motion, but it calibrates how much of the remaining gap is detection versus association versus deformation.
The key ablation removes exactly the proposal's central ingredient: the same network, same tracker, same budget, but with the deformation field restricted to a global rigid transform, so only the phase-conditioned non-rigid component differs. If this ablation matches the full model, the core claim is wrong and the paper becomes a negative result about rigid correction being sufficient.
Why not just…?
Why not just immobilize the tissue or gate image capture to the cardiac and respiratory cycle? This is what wet-lab practitioners actually do, and it fails for a documented reason and a confounding reason. The documented one: Fu et al. imaged a physically stabilized mouse kidney and still needed multi-stage non-rigid correction because breathing and pulsatility persist through restraint Four Dimensional Image Registration for Intravital Microscopy. The confounding one: clamps and suction windows compress microvessels and alter shear stress, and mechanical endothelial disruption itself triggers the Mac-1-dependent neutrophil infiltration documented after balloon-catheter injury by Welt et al. Neutrophil, Not Macrophage, Infiltration Precedes Neointimal Thickening in Balloon-Injured Arteries — the stabilization apparatus would manufacture the very signal the drug study is trying to measure. Gating additionally slashes effective frame rate below what rolling-velocity measurement requires.
Why not just band-pass/notch filter, or use standard cross-correlation registration? The spectra overlap: chest excursions and cardiac pulsation occupy the same 0.2–3.5 Hz band as the micro-events of interest, so frequency filters remove signal along with artifact, which is the core motivation stated across the motion-artifact literature, including Wital: A COTS WiFi Devices Based Vital Signs Monitoring System Using NLOS Sensing Model and Ambulatory Atrial Fibrillation Monitoring Using Wearable Photoplethysmography with Deep Learning. Cross-correlation registration fails on raw low-contrast intravital frames for reasons Fu et al. describe explicitly, and the pattern that generic features lose to physics-aware learned ones is already quantified in this domain: HOG reached an AUC of about 0.73 where artiNet's scanning-aware architecture reached 0.90 Motion Artifact Detection in Confocal Laser Endomicroscopy Images. The classical-registration baseline will still be run, so this is settled empirically rather than by assertion.
Why not just use the static post-mortem assays, since they already produce clean statistics? This is the most serious objection, because Kim et al. report highly significant separations with static endpoints — cortical Ly6G reduction at p = 0.0016, migration capacity at p = 0.0035, plasma CitH3 at p = 0.0388, brain CitH3 at p = 0.0417 Exercise preconditioning alleviates photothrombotic ischemic stroke in mice by orchestrating neutrophils. But significance is not mechanism: those assays cannot distinguish neutrophils that never rolled from ones that rolled, arrested, and detached, and Li et al.'s six-hour chemokine-window finding is invisible to any terminal assay except by killing separate cohorts at each time point, which imports inter-animal variance and precludes within-subject dose-response measurement. The proposal's decisive check on this objection is a reconciliation experiment: dynamic kinetic measurements must predict the static endpoint differences (Ly6G counts, CitH3) obtained from the same animals, and should do so with fewer animals per effect detected. If they do not reconcile, the objection wins and the dynamic assay is not measuring recruitment.
Proposed Experiments and Decisive Checks
Experiment one injects synthetic ground truth. Starting from the CTMC corpus of 86 live-cell videos CTMC and datasets from the Cell Tracking Challenge The Cell Tracking Challenge: 10 years of objective benchmarking, apply periodic non-rigid warps with physiologically calibrated amplitudes and frequencies (0.2–3 Hz, amplitudes spanning cell-scale to organ-scale) to sequences with known cell trajectories. Because true positions are known, this isolates motion-correction ability from detection ability, and yields the headline learning curve: tracking IDF1 and MOTA as a function of deformation amplitude for the proposed model versus each baseline. The Cell Tracking Challenge's silver-standard corpus, which raised cell coverage from 17.8% under gold truth to 99.1% on average The Cell Tracking Challenge: 10 years of objective benchmarking, supplies scalable reference annotations; no aggregate performance index is reported on Lacuna for it, so per-dataset reporting is used.
Experiment two is the kidney benchmark: reproduce the Fu et al. pipeline on their intravital dataset, then run both pipelines at matched compute and compare per-pixel SSD improvements against their reported 7.8% to nearly 30%, plus — the metric they could not compute — the number of immune-cell tracks maintained across all 61 time points, and the phase-invariance test on track statistics.
Experiment three is the drug assay: in the CLP sepsis cohort with Ferrostatin-1, quantitatively compare rolling-velocity distributions, firm-arrest probabilities, and diapedesis rates across untreated, pre-treated, and six-hour-post-onset groups, against the untreated 48-hour mortality of roughly 70% and the 243 co-regulated genes reported by Li et al. Ferrostatin-1 improves acute sepsis-induced cardiomyopathy via inhibiting neutrophil infiltration through impaired chemokine axis as the static anchors to reconcile with. The pre-registered success criteria are mechanistic (arrest-probability reduction with rolling velocity preserved, if the chemokine-axis account is correct) and practical (detection of the drug effect with fewer animals than a static-endpoint design of equal power). If the post-onset window closes — arrest rates unchanged when dosing is delayed — that is itself the first direct, kinetic answer to a therapeutic-timing question that every paper in this file flags as open.
References
- Remote Respiration Monitoring of Moving Person Using Radio Signals
- Four Dimensional Image Registration for Intravital Microscopy
- The Cell Tracking Challenge: 10 years of objective benchmarking
- CTMC: Cell Tracking with Mitosis Detection Dataset Challenge
- Ferrostatin-1 improves acute sepsis-induced cardiomyopathy via inhibiting neutrophil infiltration through impaired chemokine axis
- Exercise preconditioning alleviates photothrombotic ischemic stroke in mice by orchestrating neutrophils
- Contrast-Phys: Unsupervised Video-Based Remote Physiological Measurement via Spatiotemporal Contrast
- Neutrophil, Not Macrophage, Infiltration Precedes Neointimal Thickening in Balloon-Injured Arteries
- Motion Artifact Detection in Confocal Laser Endomicroscopy Images
- Tiny-PPG: A Lightweight Deep Neural Network for Real-Time Detection of Motion Artifacts in Photoplethysmogram Signals on Edge Devices
- Wital: A COTS WiFi Devices Based Vital Signs Monitoring System Using NLOS Sensing Model
- Ambulatory Atrial Fibrillation Monitoring Using Wearable Photoplethysmography with Deep Learning