Video-rate single-fiber imaging using dual optical frequency combs and deep learning

Endomicroscopy provides cellular-level visualization inside the body without physical biopsies, useful for gastrointestinal, pulmonary, and neurological diagnostics. Conventional high-resolution endoscopes rely on pixelated detector arrays (CMOS/CCD), which are too bulky for probes at the hundred-micron scale required to access confined anatomical regions such as deep brain tissue.

• The system performs a hyperspectral multiply-accumulate (MAC) across spectral channels. The 2D spatial hypercube is physically compressed into a single analog temporal electrical voltage after photodetection, a snapshot measurement that carries all spatial information in one compressed readout.
• A transformer-based deep-learning model, trained for this application, inverts the compressed snapshot to reconstruct a high-fidelity 2D image rapidly. The model learns the correlations between speckle patterns and bucket intensities and outperforms classical algorithms in both raw imaging accuracy and speed.

By combining the parallelism and stability of optical frequency combs with a tailored transformer reconstruction model, the researchers removed the sequential-projection speed limit and reduced computational latency. The optical front end is simplified to a single-core fiber and a single-pixel detector, which is attractive for probes that must be very small or disposable.

The technique directly addresses the two primary barriers to real-time single-fiber endomicroscopy: projection speed and reconstruction latency. It demonstrates a path to video-rate imaging through a zero-dimensional optical front end. The work is presented as a step toward practical deployment in biomedical applications such as single-use endomicroscopic probes, with potential relevance for any application requiring miniature, photon-efficient imaging in constrained spaces.

Concrete technical features to note

• Parallelized speckle pattern generation using dual frequency comb lines.
• Hyperspectral MAC for physical compression of spatial data into one voltage signal.
• Transformer-based deep-learning inversion tuned to the speckle-to-bucket measurement mapping.

The paper demonstrates that combining stabilized dual optical frequency combs with a purpose-built transformer model can achieve video-rate, high-fidelity imaging through a single optical fiber and single-pixel detector. This reduces hardware complexity and addresses historical speed and latency problems that have limited single-fiber ghost imaging for clinical endoscopy.