In the lab, AI rebuilds high-resolution images from faster FLIM scans

A scan that normally demands patience could yield its detail faster. In the lab, UCLA researchers used an artificial-intelligence system called FLIMPSR to reconstruct high-resolution fluorescence lifetime images from data captured at five times lower spatial resolution, without changing the imaging hardware.

FLIM — fluorescence lifetime imaging — measures how long molecules glow after a pulse of light. That duration carries biochemical information about tissue metabolism and health, without dyes or stains. The problem is that gathering enough light to measure the lifetime at every point requires slow scanning, especially over large areas such as tumor tissue.

FLIMPSR uses a generative adversarial network, a type of neural network trained to rebuild missing visual detail, to recover spatial features lost in faster scans. The UCLA team, led by Professor Aydogan Ozcan with Professor Laura Marcu’s group at UC Davis, reported a 25-fold jump in effective resolution when the system reconstructed images from pixels five times larger than usual. A diffusion model tested on the same task took more computation and produced more artifacts, while FLIMPSR delivered cleaner results in a fraction of the processing time.

The researchers tested the method on head and neck tumor tissue from patients, using high-resolution scans from three patients that had been kept out of training. The reconstructions closely followed the original images and retained fine tissue structure. The result has a clear boundary: when the resolution gap grew beyond fivefold, both the GAN and diffusion approaches began to break down. Extra training helped the system cope with noisier scans, including conditions associated with lower-cost equipment or faster acquisition.

So what, concretely? If the method also works with wide-field FLIM, it could let clinicians obtain biochemical tissue information more quickly, including during surgery where images may need to be processed in real time. That possibility remains a research target, not a deployed clinical tool: the evidence so far comes from laboratory tests on held-out tumor scans, and the UCLA group’s next step is to test the approach under different imaging conditions.