Nikon has begun reviewing information and technical materials relating to a microscopy video that won first place in its Small World in Motion competition, after microscopy researchers questioned how accurately it represented what the microscope had actually captured and where the boundary lay between enhancing a scientific image with AI and adding details unsupported by the original measurements.
Researcher Ning Xu won first place in the competition last September for a microscopy video showing the abnormal movement of cilia in a sample taken from a child with primary ciliary dyskinesia.
Cilia are hair-like structures lining the airways, while primary ciliary dyskinesia is a genetic disorder affecting their movement and respiratory function. After the result was announced, microscopy specialists raised questions about some of the details visible in the video.
They said structures in it appeared in forms and displayed behaviours that were difficult to explain biologically, while others noted that some elements seemed to appear within the scene and then disappear. These questions do not prove that the video was generated by AI.
Ning Xu explains his use of AI
Xu said the technology was not used to generate the original video or create the cilia or their movement. Instead, it was introduced at a later stage to process grey-scale microscopy data, highlight similar structures and colour them, with the aim of improving their visual presentation. The case is connected to a broader shift in scientific and medical imaging.
Whereas the process once consisted mainly of the microscope capturing what was in front of its lens and the image then being processed to improve its clarity and readability, algorithms can now reduce noise and enhance details, as well as reconstruct parts of images and infer what might be present in them.
The issue is not the use of AI itself, as algorithms have become important tools in the field and can be used to reduce noise, increase resolution, correct motion and reconstruct images from incomplete data.
The dispute, however, concerns whether processing reveals information already present in the original data or produces details that were not recorded by the initial measurement. When an algorithm enhances a low-resolution image, the final result may appear far clearer than the original, but greater clarity does not necessarily mean that the microscope actually captured every detail visible in the resulting image.
The issue becomes more complex when an algorithm can produce structures that appear convincing to the human eye despite not being supported by sufficient experimental data. Recent research in medical imaging has warned of this kind of “visual hallucination”, as image-restoration algorithms can add structures that were not present in the original or remove genuine ones, without the difference being easy for a human observer to detect.
This puts the reliability of the final result to a test that goes beyond the quality of its presentation. The standard differs from the treatment of artistic images, where changing colours or adding elements may be part of the nature of the work.
The limits of acceptable processing in scientific images
A scientific image, by contrast, is linked to data that are expected to represent a phenomenon or experiment that can be examined and independently verified. This makes the integrity of the original data and disclosure of the processing steps essential to its assessment. In this context, recent scientific guidelines stress the protection of original data and the clear description of the digital processing used.
Medical publishing guidelines, including those updated by the Journal of the American Medical Association, allow standardised adjustments to brightness, contrast and colour, but stress that such adjustments must not highlight, conceal or alter the meaning of scientific elements. The guidelines also call for disclosure of the tools used in processing and for original images to be made available for review when necessary.
Researchers in the British scientific journal Nature Methods have also drawn up guidance aimed at improving the clarity and reproducibility of microscopy images, including the documentation of image-processing operations and analytical methods. The review of the video is particularly sensitive because Nikon’s competition rules prohibit entries generated by AI and give the company the right to request the original work for verification.
Nikon re-examines the video’s technical materials
After the questions were raised, Nikon said it was re-examining the information Xu provided during the initial judging process, along with additional technical materials he later submitted. The additional materials include details of the imaging equipment he used, the methods he followed to obtain the images and how he processed them.
The question therefore centres not simply on the use of AI, which Xu acknowledged during the processing stage, but on the nature of the changes the algorithms made to the data after it was captured.
The case tests the traceability of scientific images
The case goes beyond the boundaries of a microscopy competition. The expanding use of AI in scientific and medical imaging requires images to be viewed as the product of a complete chain that begins with raw data, passes through processing and analysis, and ends with the final version shown to researchers or the public.
This may prompt researchers, scientific journals and competitions to request more detailed information about the type of algorithm used, what it changed within the image, which elements remained unchanged and whether the final result can be compared with the original data. Trust will then depend on whether every visible detail can be traced to a real measurement, rather than simply on whether the image looks realistic and convincing.
The challenge algorithms pose to microscopy lies in their ability to produce highly convincing images, making it harder to distinguish what the lens recorded from what the AI inferred. Nikon’s review of the information and technical materials submitted by Xu remains a practical test of how these boundaries should be applied to digitally processed scientific images.
