Endomicroscopy provides critical pathophysiological information for the diagnosis of gastrointestinal, pulmonary, and neurological abnormalities by enabling in vivo, cellular-level visualization of tissues without the need for physical biopsies. However, traditional high-resolution endoscopes face severe physical limitations; bulky pixelated detector arrays (like CMOS or CCD) cannot be miniaturized to the hundred-micron scale required to access highly constrained anatomical regions, such as the deep brain, without compromising resolution.
While optical-fiber-based ghost imaging (also referred to as single-pixel imaging) is theoretically ideal for these photon-starved, size-constrained environments, its real-world adoption has been limited by several technical challenges, including a fundamental speed bottleneck. Conventional systems require slow, sequential pattern projection using spatial light modulators or wavelength sweeps, which limits frame rates and introduces severe motion artifacts when imaging dynamic objects. Furthermore, classical reconstruction algorithms suffer from substantial computational latency, making real-time, video-rate image reconstruction impractical.
In a new paper published in Light: Science & Applications, a collaborative team of scientists from NTT Research, the University of California Irvine, and Los Alamos National Laboratory introduces a groundbreaking hardware-software co-design that completely resolves these historical bottlenecks. The research demonstrates a novel paradigm in computational imaging that achieves video-rate capture of dynamic scenes while utilizing a spatially non-resolving "zero-dimensional" optical front end consisting exclusively of a single-core optical fiber and a single-pixel photodetector.
To eliminate the sequential projection bottleneck, the researchers leveraged highly stabilized dual optical frequency combs, mapping individual comb lines to mutually uncorrelated speckle patterns. This enables fundamentally parallel pattern generation and detection. The system performs a parallelized, "hyperspectral" multiply-accumulate (MAC) reduction across the spectral channels, physically compressing the hypercube containing 2D spatial information into a one-dimensional analog temporal electrical voltage signal after photodetection.
To invert these highly compressed snapshot measurements, the team developed an application-specific deep-learning transformer model. This tailored algorithm deeply learns the complex correlations between speckle patterns and the corresponding bucket intensities of the target-encoded speckle pattern, reconstructing a high-fidelity target image, outperforming classical algorithms in both raw imaging accuracy and processing speed.
"By merging the inherent parallelism and precision of optical frequency combs with the remarkable reconstruction capabilities of a transformer-based deep learning model, we have eliminated the sequential projection bottleneck and significantly improved imaging speed and fidelity," said Dr. Myoung-Gyun Suh, Senior Scientist and Group Head at NTT Research's Physics and Informatics Laboratories and lead author of the study. "Our hardware-software co-design dramatically simplifies the optical front end, making high-fidelity dynamic ghost imaging highly viable for size-constrained scenarios like single-use endomicroscopic probes".
This innovation marks a significant leap toward the practical deployment of advanced ghost imaging technologies in real-world biomedical applications, freeing the technique from the optical table and moving it closer to clinical neurosurgical and diagnostic use.
