A nanophotonic encoder has reconstructed visible hyperspectral images from light carried through one optical fiber, covering wavelengths from 400 to 700 nanometres and incident angles from minus 60 to plus 60 degrees. The Nature Communications study matters because it moves the spatial and colour encoding to the probe tip, while the fiber transports only a spectrum that remains stable when the cable bends or twists.
Hyperspectral imaging records a spectrum at every point in a scene. Conventional instruments usually separate those spatial and spectral measurements with cameras, scanning optics or fiber bundles, so the probe gets wider as it collects more points. A multimode fiber can carry many spatial patterns, but small bends scramble those patterns. The new system avoids reading the pattern at the far end. It instead preserves information in the total optical power measured at each wavelength.
The component at the fiber tip is a disordered-dispersion encoder, or DDE. It combines a metagrating with a multilayer thin film so that each incoming angle and wavelength receives a distinct transmission signature. A spectrometer records the mixed spectrum after it crosses the fiber, and a sparsity-constrained algorithm solves backward for the original spectral-angular distribution. In the prototype, the encoder and a polarizer sat beside the end face of a 400-micrometre-core multimode fiber.
Tests with one-dimensional targets returned average errors of 6.44% across the angular range and 6.27% across the spectrum. Those figures describe reconstruction against the experiment's own ground truth, rather than a matched comparison with commercial probes. Simpler periodic-grating and thin-film encoders used inside the study did not recover the complete spectral-angular signal. In a separate robustness test, 91 measurements taken while the fiber was bent or twisted by hand showed little change in the spectral signal, supporting the decision to transmit wavelength totals instead of fragile spatial modes.
Two-dimensional images still required repeated views. The team rotated a four-colour windmill target and obtained regional spectral errors from 2.56% to 8.30% against a commercial hyperspectral camera. An ex vivo mouse gastric-tumour tissue slice was then measured at 36 rotation angles, with three independent runs averaged before spectral clustering separated three stained regions. The colours tracked differences in staining and morphology, but this was a proof of concept on one prepared slice, not a diagnostic or clinical validation.
The result establishes a compact encoding route, not a finished endoscope. The demonstrated two-dimensional images need rotation because the present DDE mainly resolves one angular axis, and the iterative reconstruction takes minutes. Supplementary tests measured an encoding response through a 2.5-micrometre single-mode fiber, but image reconstruction there was simulated and required a high-power laser to offset the low photon count. The public code likewise uses simulated objects and measurements built from calibrated response matrices.
