Pixel super-resolved fluorescence lifetime imaging using deep neural networks.
Image Credit: Ozcan Lab @UCLA
UCLA researchers developed a deep learning framework that reconstructs high-resolution fluorescence lifetime images from data acquired at up to five times lower spatial resolution, offering a path toward faster tissue imaging without complex hardware.
Fluorescence lifetime imaging, or FLIM, lets researchers see biochemical detail in tissue without dyes or stains, by tracking how long molecules glow after being hit with a pulse of light. That glow duration carries information about metabolism and tissue health that standard microscopy cannot capture, which is why FLIM has attracted interest for applications such as cancer diagnostics and surgical guidance. The catch is time: collecting enough light to accurately measure lifetime at every point in an image means scanning slowly, especially across large tissue areas.
A team at the UCLA Samueli School of Engineering set out to work around that limit. Led by Professor Aydogan Ozcan and in collaboration with Professor Laura Marcu’s group at UC Davis, the researchers built a deep learning system called FLIMPSR that reconstructs higher-resolution FLIM images from scans captured at much lower resolution. Paloma Casteleiro Costa, a postdoctoral scholar in the UCLA Department of Electrical and Computer Engineering, is the first author of the study, now published in PhotoniX, a Springer Nature journal.
At its core, FLIMPSR is a generative adversarial network (GAN) trained to reconstruct spatial features lost when FLIM data is collected quickly. The team showed the model could recover accurate images from scans taken with pixels five times larger than usual, a 25-fold jump in effective resolution, without any change to the imaging hardware itself. They also tested a diffusion model, a newer and more computationally intensive class of generative AI, on the same task, and found their approach produced cleaner results in a fraction of the processing time, while being less likely to introduce artifacts that did not reflect the real tissue, a concern for any tool that might eventually inform a medical decision.
To test the system, the researchers scanned head and neck tumor tissue from patients and set aside three patients’ worth of high-resolution scans that the model never saw during training. Applied to that held-out data, FLIMPSR reconstructed images that closely tracked the original high-resolution scans, preserving fine tissue structure that would normally be lost at faster scan speeds. Reconstruction quality held up across a fivefold resolution gap; pushing much further, both the GAN and the diffusion model began to break down.
The researchers also demonstrated that a short round of additional training can make the model hold up under noisier scanning conditions, the kind that might come from lower-cost equipment or more aggressive scan speeds, without hurting its performance on cleaner data.
Next, the research group plans to explore whether the same approach extends to wide-field FLIM systems, which face a different version of the same resolution-speed trade-off. If the method holds up under those conditions, it could help move high-resolution FLIM out of the research lab and into faster-paced clinical settings, including surgery, where images need to be processed in real time.
See the article:
Paloma Casteleiro Costa, Parnian Ghapandar Kashani, Xuhui Liu, Alexander Chen, Ary Portes, Julien Bec, Laura Marcu, and Aydogan Ozcan, “Pixel super-resolved fluorescence lifetime imaging using deep neural networks”, PhotoniX (2026)
https://link.springer.com/article/10.1186/s43074-026-00277-9