Implicit neural image field for biological microscopy image compression

Microscopy datasets have grown large enough that storing, sharing and viewing them has become a practical constraint. General-purpose image and video compression was designed for photographs and film, and handles poorly the images microscopes produce, which can span three spatial dimensions, several fluorescence channels and time. This practice stores an image differently: a small neural network is trained to reproduce that one image from the coordinates of its pixels, and what gets saved is the network rather than the picture. The size of the network is fixed in advance, so the degree of compression is set by the user rather than discovered afterwards. Recovering the image means querying the network at chosen coordinates, which allows a single slice, an irregular region or a quick low-resolution preview to be retrieved without unpacking the whole dataset.

The work joins a computational imaging group at Peking University with the AMBIOM groups at a Leibniz institute (ISAS) in Dortmund and an experimental immunology group at University Duisburg Essen. The methodology and the main experiments came from the computational side; the Dortmund and Essen groups contributed data collection and curation and guidance on the biomedical application. Two elements made the approach practical. Earlier attempts to compress images this way required training settings to be tuned separately for each image or image block, which takes considerable time and becomes especially pronounced for large bioimages. Here the training is driven by a general-purpose optimiser that had itself been trained on large numbers of optimisation problems and was then adapted for compression, removing that tuning step. Second, compression can be steered toward the use the data is intended for: an additional objective can be added so that the stored image preserves whatever a downstream segmentation depends on, or so that noise in images taken at low illumination is suppressed instead of faithfully reproduced. All images used for development and testing were taken from public repositories. The work relies on GPU computing.

The approach was tested on three-dimensional cell images, multichannel tissue images and five-dimensional time-lapse recordings, at target compression ratios between roughly one hundred and five hundred fold. It reached the compression ratio requested, which the conventional video codec frequently did not. On the three-dimensional stem cell images, measured across twenty-five samples at three compression ratios, it scored higher on two standard image-fidelity measures than both the codec and the earlier neural representation approach. Guided compression preserved the fine striped structure needed for a downstream structural analysis, and in images taken at very low laser power the codec produced no usable reconstruction while the guided method retained the visible structures.

The cost is that compressing an image means training a network: minutes per sample against seconds for the conventional codec, and slower than codecs on the same hardware generally. A hybrid mode reduces this by letting a fast codec do most of the work and training the network only on the difference. The code is released under a permissive open licence with a persistent identifier, the exact image files used are deposited separately with their own identifier, and the paper is open access.

Dai, G., Zhang, R., Wuwu, Q., Tseng, C.C., Zhou, Y., Wang, S., Qian, S., Lu, M., Tuz, A.A., Gunzer, M. and Huang, T., 2025. Implicit neural image field for biological microscopy image compression. Nature Computational Science, 5(11), pp.1041-1050. https://doi.org/10.1038/s43588-025-00889-4