Simple Future for Cell Microscopy: AI Generates Multiple Biological Views From One Hologram

A microscopy system that needs only one inexpensive holographic image could recover detailed information about a cell's physical structure and also predict where its nucleus, mitochondria and membrane would appear in fluorescence images, according to a new study by Seonghwan Park and colleagues. The research, 'AI-driven dual-mode phase and label-free fluorescence imaging platform using a single-shot Gabor hologram,' was published in nature journal Microsystems & Nanoengineering in 2026.

The researchers developed an artificial intelligence platform that turns a single-shot Gabor hologram into both a quantitative phase image and three virtual fluorescence channels, reducing the need for fluorescent staining, repeated image acquisition and complicated multimodal optical hardware during actual use. The approach combines holographic imaging with a one-sided unsupervised diffusion model, allowing the system to learn from phase and fluorescence reference images without requiring every hologram to be directly paired pixel-by-pixel with those targets.

One simple hologram replaces much of the optical complexity.

Fluorescence microscopy is valuable because fluorescent markers can reveal specific biological structures and molecular activity, but staining takes time, alters experimental workflows and can cause photobleaching or phototoxicity that becomes particularly troublesome when researchers want to observe living cells repeatedly. Quantitative phase imaging avoids labels and captures properties such as optical path difference, dry mass and morphological changes; it does not naturally provide the organelle-specific biochemical contrast supplied by fluorescence imaging.

Traditional systems that combine both forms of imaging usually need several light sources, detectors, synchronized optical pathways and fluorescent labels. Gabor in-line holography offers a much simpler alternative because a compact arrangement of a coherent light source, beam expander, microscope objective and camera can record an interference pattern in one shot; these holograms are affected by twin-image artifacts and limited contrast.

The new platform pushes much of that complexity from hardware into computation. During training, the researchers used a conventional dual-mode system to obtain accurately aligned phase and fluorescence reference images, including fluorescence channels for nuclei, mitochondria and cell membranes. Once training was complete, that elaborate imaging arrangement was no longer needed for inference; the deployed model worked only from a single Gabor hologram.

A one-way diffusion model learns without directly paired hologram targets.

At the heart of the system is a one-sided unsupervised diffusion framework designed around the fact that holograms and the dual-mode reference images were collected under different acquisition conditions and were therefore not directly paired. Instead of learning a two-way mapping similar to cycle-consistent systems, the model concentrates on the useful direction: turning a hologram into phase and fluorescence information.

Training happens in two stages: A first network generates synthetic Gabor holograms from dual-mode images using patch-level contrastive learning and adversarial supervision, creating a bridge between the otherwise unpaired datasets. A second hologram-conditioned diffusion network progressively denoises an image while using the hologram as guidance, reconstructing both phase and virtual fluorescence information. The authors argue that dropping cycle-consistency reduces unnecessary computational work because reconstructing a physically meaningful hologram from phase and fluorescence images is not required during deployment.

Inference is simpler than training: The reconstruction network starts from random noise and refines it over four denoising steps, conditioned on the measured hologram, ultimately producing the quantitative phase image along with virtual fluorescence outputs for the nucleus, mitochondria and membrane.

The study used three human cancer cell lines: SNU-475 liver cancer cells, T-24 bladder cancer cells and NCI-H1299 non-small-cell lung cancer cells, and examined untreated cells alongside cells exposed to FCCP, Rotenone, Staurosporine or Blebbistatin. The main dataset contained 750 dual-mode samples across 15 experimental conditions, with 600 allocated to training and 150 to testing before augmentation.

The reconstructed images preserved features useful for cell and drug analysis.

Across its experimental validation, the platform achieved an overall FID of 57.74, SSIM of 0.76, PSNR of 26.89 dB and LPIPS of 0.12, while recovering both cellular morphology and organelle distributions from the single holographic input.

Training on phase and fluorescence information together produced clearer results than training separate models on only one modality. For fluorescence reconstruction, dual-mode learning reduced FID from 80.67 to 52.20 for nuclei, from 80.30 to 64.31 for mitochondria and from 105.20 to 72.74 for membranes, while also improving SSIM and LPIPS across most fluorescence channels.

The generated images were useful beyond visual comparison: Measurements including optical path difference, dry mass, projected area and circularity closely reflected those obtained from ground-truth images, while fluorescence-guided segmentation helped define cell boundaries and organelle-specific regions.

When these morphological features were used to distinguish lung, liver and bladder cancer cells, the combined feature set produced an overall classification accuracy of 75%, with F1 scores of 0.842 for lung, 0.758 for liver and 0.651 for bladder cells. Mitochondrial and membrane features generally proved more discriminative than nuclear features.

Drug-related morphological changes were also retained: Rotenone-treated lung cells showed elongation and irregular shape, liver cells displayed strong changes following Rotenone and Staurosporine exposure, and bladder cells showed subtler alterations because of their naturally higher morphological variability. Quantitative analysis detected significant treatment-related changes in optical path difference, dry mass, projected area and circularity.

Five-class drug-response classification reached 47% for lung cells, 47% for liver cells and 42% for bladder cells, well above the 20% chance level, with membrane-derived features producing the most consistently useful discrimination.

Robust performance points toward cheaper and more flexible bioimaging

The model was tested under several conditions it had not seen during standard training. A network trained using 20× images still reconstructed images collected at 40× magnification, and reducing illumination to 50% and 33% of normal power caused little change in SSIM or LPIPS. Predictions also remained comparatively stable when the hologram was deliberately defocused between −3 and +3 micrometres.

Human iPSC-derived cardiomyocytes produced weaker reconstructions when processed by the cancer-cell-only model, particularly in fluorescence channels, although adding just 80 cardiomyocyte images to training brought a small but consistent improvement. The authors identify training-data diversity as an important remaining limitation.

The system processed each image in about 89.47 milliseconds, or roughly 11 images per second, using 96.97 million parameters and about 1.09 GB of GPU memory, also cutting parameter count and computational operations substantially compared with a bidirectional diffusion baseline.

The strongest practical promise of the work lies in separating training complexity from everyday imaging. Fluorescence labels and the conventional dual-mode microscope are still needed to create reference data during model development, so the method does not remove fluorescence microscopy from the training process. Once trained for an appropriate biological setting, a much simpler holographic setup can potentially provide phase measurements and virtual fluorescence information without staining each new sample, opening a path toward repeated live-cell observation with less concern about photobleaching and phototoxicity.

The researchers see future extensions involving other imaging contrasts, live-cell temporal information and automated analytical tools, suggesting that the broader value of the approach may be its ability to shift sophisticated microscopy from increasingly complicated optics toward intelligent computational reconstruction.