Nikon’s Microscopy Disqualification and the Limits of AI Post-Processing

The ruling was about eligibility

Nikon revoked the first-place award in its 2026 Small World in Motion competition after a second review of Ning Xu's microscopy video. The entrant had described the AI work as post-processing of reconstructed experimental data. Nikon concluded that the entry breached its generative-AI rules.

PetaPixel's October 9 report quotes the prohibition, “AI-generated videos are not permitted,” and Nikon's account of reviewing the entry and supporting materials with the judging panel. Nikon also said its decision concerned competition eligibility, not the entrant's professional reputation, contributions, or intent. That scope matters: disqualification does not establish a finding of research misconduct.

Nikon did not publish the model or complete processing sequence used for the entry. The entrant’s account and the researchers’ objections instead point to the stages between acquisition, reconstruction, and the colored output. Those stages determine which shapes in the finished movie can be traced to the recorded data.

What happened between the award and the review

The winning video was presented as showing cilia moving in airway tissue obtained from a child with primary ciliary dyskinesia. Cilia are fine structures on cell surfaces. Beneath them, the finished video displayed strongly colored forms whose identities and behavior attracted questions from microscopy researchers.

CNN's October 3 account places the award in mid-September and describes the entrant's response. Xu said AI had not generated the experimental video, the cilia, or their motion. He described applying it after grayscale reconstruction to distinguish and color structures with similar morphology.

By PetaPixel's October 1 report, Nikon was reassessing information from the original review together with additional technical documentation. The October 9 report then announced disqualification. These are dates of contemporary reports, not an inferred schedule of private deliberations.

The interval included a substantive disagreement. Researcher Edward Phelps questioned purple structures appearing and disappearing, as well as the appearance and scale of green cilia-like features. Those objections concerned whether the scene matched the biology it appeared to depict. Xu's response concerned the purpose and sequence of his processing. A description of purpose does not, on its own, demonstrate that the resulting boundaries and structures match the recorded data.

“Post-processing” names a stage, not a guarantee

Post-processing happens after acquisition. That includes modest display adjustments, numerical filtering, reconstruction, segmentation, and operations that infer missing detail. The label tells a reader when something happened but does not specify what the algorithm changed.

Coloring illustrates the ambiguity. Assigning a display color to an already defined channel is different from asking a model to identify and separate structures before coloring them. In the latter case, the colored result inherits any mistakes in the separation. If two adjacent objects are treated as one, changing their color does not recover the missing boundary.

This is why the researchers' questions can remain relevant even if the entrant accurately describes the operation as enhancement. The unresolved technical issue is what the intermediate data showed and what survived each transformation. Nikon's quoted prohibition and final decision establish the competition outcome; they do not supply a general catalogue of which scientific image-processing methods are acceptable.

Public discussion under Nikon's review notice included requests for original images and processing details. Researcher Dylan Burnette also described experimenting with AI on his own microscopy material. His example was not a reconstruction of Xu's workflow. It illustrated how plausible an altered image can look, rather than proving which operation produced this entry.

Three kinds of processing change images differently

A brightness or contrast adjustment changes how measured values are displayed. The image guidance hosted by the US Office of Research Integrity distinguishes moderate adjustments across an image from transformations that lose information. If different bright values are all clipped to white, their differences disappear from the displayed result. No generative model is required to create that loss.

Nonlinear adjustments can change relationships between intensities too. That matters when the intended result is a measurement of brightness rather than a picture for orientation. The method and its scope explain what a reader can infer from the processed image more precisely than the word “enhanced.” This educational guidance is separate from Nikon's competition rules.

A conventional local filter combines neighboring pixel values. It can suppress fluctuations, but a small real structure can also resemble the fluctuation the filter is designed to remove. The ORI-hosted filtering guidance discusses changes to pixel values and possible artifacts, along with documenting the software and operation. Traditional processing is not automatically harmless.

Learned restoration uses relationships acquired from training data to predict a cleaner output. The Janelia bioimaging AI guide explains supervised restoration using noisy images paired with cleaner reference images. Such a model can make weak signals easier to interpret, but it does not possess a direct measurement of the true structure for every new input.

The intended measurement determines which errors matter. For example, an extra apparent object changes a count, and a shifted edge changes a size estimate. In a movie, inconsistent reconstruction across frames can resemble motion. Preserving counts and positions is therefore part of evaluating a restoration for those measurements.

For intensity comparisons, the ORI-hosted measurement guidance generally favors raw data to avoid processing artifacts. When calibration or normalization is needed, it calls for uniform treatment across the data and documented procedures. Increasing contrast in one specimen alone can mix a display difference with the biological difference the measurement is meant to capture.

A published example of a mismatched model

A 2021 Nature Methods article on reproducibility in deep-learning bioimage analysis provides a concrete comparison in Figure 2a. This independent experiment restores spinning-disk confocal images of cell nuclei. A CARE model trained on suitable nuclei data reduced the noise, while a model trained on structured illumination images of F-actin failed to restore the same input correctly. Holding the input image constant makes the effect of the different training material visible.

The Janelia guide describes comparisons with appropriate reference images and discusses limitations of restoration. A higher-quality view of the same structures can expose changes in location or shape. An average image-quality score may underweight a small region that matters to the eventual biological measurement, so inspecting the relevant objects adds information that the global score can miss.

Why intermediate images matter as much as the original

Acquisition, reconstruction, separation of structures, coloring, and final rendering can each change the result. Comparing only the first and last image can reveal a difference without identifying the step that introduced it. If grayscale reconstruction precedes AI coloring, as Xu described, the reconstructed grayscale output is a useful comparison point in its own right.

The acquisition data show what the instrument recorded. The reconstruction shows how those measurements became an image. A subsequent labeled or colored output reveals whether a boundary was added, two structures were merged, or one structure was split. In a movie, matching the same location across frames also helps distinguish continuity in the recorded scene from changes introduced during processing.

The Nature Methods article calls for reporting methods, models, data, and validation in ways that support reproducibility. Naming a tool alone leaves out the training material, settings, and intermediate operations needed to understand a result. A connected record of these stages allows someone else to examine where a discrepancy first appeared.

Nikon named Nguyen Nam Nhat's video of a roundworm and the single-celled organism Dileptus as the new winner. The organizer also said the incident had prompted it to revisit rules and evaluation practices for future entries.