Blur-Tolerant Fly-Through Scanning as the Readout for Cytokinin–Auxin Screens

Plant tissue culture works by placing a piece of plant tissue on a gel containing two hormone families. Cytokinins such as 6-benzylaminopurine (6-BA) push cells toward making shoots. Auxins such as naphthaleneacetic acid (NAA) and indole-3-butyric acid (IBA) push them toward making roots. Intermediate ratios produce callus, an undifferentiated cell mass. Finding the right ratio for a new species is still done by brute force. A lab runs an orthogonal grid of concentrations, waits four weeks, and then destructively counts shoots or roots. The species-specific optima in Illicium henryi micropropagation (1.0 mg/L 6-BA with 0.05 mg/L NAA) and Quercus suber tissue culture (0.60 mg/L 6-BA for shoots, IBA plus NAA for rooting) were found this way. The dose response is non-linear: in Achyranthes bidentata, moderate NAA plus 6-BA helped growth and the highest dose suppressed it. A coarse grid read once at the end therefore wastes vessels and weeks.

We propose to replace the end-point count with frequent, non-destructive 3D scans of every vessel, taken by a camera that sweeps continuously along a shelf without stopping. Then we test whether early 3D growth trajectories rank hormone treatments the same way the four-week destructive count does. Continuous sweeping causes motion blur. The camera cannot simply use a shorter exposure, because the light budget is constrained: shorter exposure needs more light, and light itself changes plant development. ChronoRoot showed photoperiod alone shifts root architecture. So the imaging can perturb the very hormone experiment it is measuring. The machine-learning contribution is a blur-aware 3D reconstruction tailored to this setting, evaluated by whether it gets the biology decision right rather than by image fidelity alone.

Core Claim

In a gantry-mounted culture scanner, the main obstacles that make blur-robust reconstruction hard elsewhere largely disappear, and the remaining ones are specific and tractable.

Plant growth during a single exposure, or even a minute-long sweep, is negligible. The static-scene assumption of BAD-Gaussians therefore holds within a scan. Growth is modeled only across scans, hours apart. This means we do not need the dynamic-object machinery of BARD-GS.

The camera trajectory is known up to small error from the motor encoders. It can serve as a strong prior on the blur path instead of relying on structure-from-motion feature matching, which tends to fail on low-texture agar.

What remains genuinely unsolved is refraction through the vessel wall and the thin, translucent structure of microshoots and young roots. Existing deblurring methods are evaluated in free air on macroscopic objects.

Our claim is that encoder-anchored, refraction-aware blur modeling lets the sweep run several times faster at a fixed light dose. It should do so without degrading the treatment ranking compared with sharp stop-and-go capture. Early-week 3D traits from these scans should predict the week-four multiplication or rooting rate well enough to run an adaptive search over the hormone grid with fewer vessels.

Research Program

Scanner and culture system. We use square polycarbonate culture vessels with flat faces. Refraction through a flat wall is then an analytic, per-ray offset given by Snell's law, not a learned field. Curved tubes and agitated liquid bioreactors are deliberately excluded. Moving liquid would break the within-scan static assumption, and we prefer to state that boundary rather than claim it away.

A single RGB camera rides a linear gantry past a shelf of vessels at controlled speed and exposure. Lighting is held at the culture room's normal photosynthetic level. The biological systems are:

  • an I. henryi shoot-proliferation grid over 6-BA and NAA on MS medium, following Li et al.;
  • a Q. suber rooting grid over IBA and NAA, following Yu et al..

Both have published optima we can check our search against.

Reconstruction. We build on BAD-Gaussians, which represents the scene as 3D Gaussian splats (small colored ellipsoids rasterized to images). It explains each blurry frame as the average of sharp renders along a camera path, with the path parameterized by a cubic spline and optimized jointly with the scene. We make two changes.

First, rays are bent at the vessel wall before entering the splat scene, using the known wall thickness and refractive index. The wall pose is fit once per vessel.

Second, the spline is tied to the encoder trajectory. For a blurry frame $B_i$ with exposure window $[t_i, t_i+\Delta t]$, the loss is

$$ \mathcal{L} = \sum_i \Big\| B_i - \frac{1}{K}\sum_{k=1}^{K} R\big(G, \rho(T(\tau_{ik}))\big) \Big\|_1 + \lambda \sum_i \int_{t_i}^{t_i+\Delta t} d\big(T(t), E(t)\big)^2 \, dt . $$

The symbols are defined as follows:

  • $R$ renders the Gaussian scene $G$ from a camera pose;
  • $\rho$ applies the wall refraction to that camera's rays;
  • $T(t)$ is the optimized spline pose;
  • $\tau_{ik}$ are $K$ sample times inside the exposure;
  • $E(t)$ is the encoder pose;
  • $d$ is a distance on rigid-body poses;
  • $\lambda$ sets how much the optimizer may deviate from the encoders to absorb vibration.

Trait extraction and growth model. From each scan we extract three traits: shoot count, total plant volume, and shoot-tip positions, the last by clustering Gaussians and skeletonizing in 3D. Consecutive scans are linked only through trait time series, not a shared 4D field.

We chose this over adopting GROWFLOW, which reconstructs growing plants with a neural ODE flow on Gaussians. That method assumes growth is monotonic, and cultures under high hormone doses brown and shed tissue: browning reached 58.89% of Q. suber explants without antioxidant in Yu et al..

Adaptive screening. A Gaussian process maps hormone concentrations and early traits to the predicted week-four outcome. It chooses which concentrations to plate in the next batch. This is compared with the orthogonal-array design at equal vessel counts.

Baselines

The naive reference is standard 3D Gaussian splatting trained directly on the blurry sweep frames with encoder poses. It shows how much blur actually costs if ignored. It uses the same frames, the same iteration budget, and the same trait extractor.

The strongest existing method is BAD-Gaussians run as published, with poses initialized by structure-from-motion. It is the strongest competitor because it already models continuous camera blur in splats and matches our static-within-scan regime. On its home benchmark, the Deblur-NeRF synthetic set, it reports a 3.6 dB PSNR gain over the previous best. BAD-NeRF reports 30.94 PSNR there. We found no reported result on Lacuna for any deblurring method on in-vitro culture vessels.

We run BAD-Gaussians on identical sweeps with the same Gaussian count, iteration budget and per-method hyperparameter search budget. We beat it if two things hold:

  • our rendering PSNR and trait error against the sharp reference scans are higher at the same sweep speed;
  • the treatment ranking it produces agrees less with the destructive endpoint than ours.

The sharp-capture reference is stop-and-go acquisition at each viewpoint. One variant uses a short exposure with brief flash illumination, the other a long exposure with the camera stopped. This is not a competitor to beat on quality. It is the ceiling, and it costs time per vessel, which we report.

The 2D reference is a top-and-side-view segmentation pipeline in the style of ChronoRoot, applied to the same vessels. It measures what the third dimension adds to ranking.

Two ablations isolate the key ingredients: removing the encoder prior ($\lambda = 0$), and removing the wall refraction model. A third ablation, plain splatting on 2D-deblurred frames, tests whether physical blur modeling matters at all here.

Why not just…?

Why not just stop the gantry, or fire a flash at each view? This is the real threat. A low-duty-cycle flash gives a sharp frame with a small total light dose, so the photobiology argument weakens. The cost is throughput, and possibly flash-induced stress, which we have not seen quantified. We therefore treat the decision as an empirical tradeoff rather than argue it.

We measure seconds per vessel for stop-and-go against continuous sweeps at three or four speeds. In parallel, we run a flash-exposed versus unexposed control grid to see whether repeated flashes shift the multiplication rate. If flash stop-and-go matches our ranking accuracy at acceptable throughput and leaves growth unaffected, the reconstruction method is unnecessary for this application. The trait-forecasting and adaptive-screening parts of the work then still stand on sharp data.

Why not just image in 2D, as ChronoRoot and RhizoVision do? RhizoVision Crown reaches R² = 0.99 on calibrated wire diameters but needed heavy edge smoothing to avoid spurious lateral roots. ChronoRoot works on roots constrained to a flat plate. Microshoot clusters at multiplication rates near 5.67 overlap heavily in projection. Still, ranking treatments may not need exact counts. The 2D baseline's rank agreement with the endpoint is the direct test. If it equals the 3D rank agreement, we will say so.

Why not just deblur each frame in 2D and then reconstruct? Deblur-NeRF reports 28.77 PSNR against 23.78 for a plain radiance field on blurry input. It argues that per-frame restoration is not consistent across views. Thin translucent shoots against agar are exactly where hallucinated detail would corrupt counts. The 2D-deblur ablation measures this directly on our vessels.

Why not just keep destructive end-point counting? For a single fixed grid it works. Its weakness is that it gives one reading per vessel after four weeks, which rules out adaptive allocation. The adaptive-versus-orthogonal comparison at equal vessel budget is what justifies non-destructive scanning at all.

Proposed Experiments

The first experiment is a fidelity check. Each vessel is scanned twice in immediate succession: once by flash stop-and-go (the reference) and once by continuous sweep at increasing speeds. We report rendering PSNR and SSIM against the reference renders, and shoot-count and volume error against hand counts. Results are reported as curves over the product of sweep speed and exposure, which sets the blur length. The prediction is that our method's curve stays flat to a blur length several times longer than BAD-Gaussians with structure-from-motion initialization. The refraction ablation should show its largest gap at oblique viewing angles.

The second experiment is the decisive biological check. We run the I. henryi 6-BA by NAA grid and the Q. suber IBA by NAA grid with scans every two to three days. At day 7 and day 14 we compute Kendall rank correlation between the predicted and the actual day-28 treatment ranking, with the day-28 ranking measured destructively. We compare all baselines on this measure. We also check whether the search recovers the published optima.

The third experiment is the vessel-budget comparison. The adaptive design and the orthogonal array each get the same number of vessels. We measure the regret of the chosen concentration against the best observed outcome. The light-dose control from the flash experiment is reported alongside, so readers can judge whether any imaging regime biased the hormone response.