Lensless Sensor Chip Records Gigapixel Video at 30 fps

A single image sensor with no lens in the optical path has resolved 308-nm linewidths across centimeter-scale fields at 30 frames per second — a throughput of 20.8 billion pixels per second from one detector. Researchers at the University of Connecticut, Storrs, led by professor Guoan Zheng and graduate student Qianhao Zhao, published the result in Nature Communications (vol. 17, article 8149, 2026). Each reconstructed frame carries 30,480 × 22,800 pixels.

The demonstration targets a measurement problem that has stood since instrument-grade optics existed: a scene — a melting snowflake, a dividing cell — that spans millimeters but carries sub-micron detail and changes from second to second. A high-NA objective resolves the fine branches over a small patch; a wide-field lens records the whole object but blurs the detail away.

What limits the pixel count?

The space-bandwidth product (SBP), formalized by Lohmann and colleagues in 1996, counts the resolvable spots a system delivers, and two ceilings cap it. Lens aberrations grow with aperture and field angle, so resolution and field of view trade against each other. The detector cannot report more resolvable spots than its total pixel count.

Existing instruments work around the ceilings rather than lift them. Whole-slide scanners and Fourier ptychography build gigapixel datasets from sequential measurements that take minutes, as Zheng, Horstmeyer and Yang showed in 2013. Camera arrays, in Brady's 2012 design, capture gigapixel snapshots in parallel by multiplying lenses and sensors. Either way, doubling the SBP doubles the measurements or the hardware.

How does the chip replace the lens?

The UConn device lifts both ceilings at once. A laser diode illuminates the sample through a fold mirror. A conventional image sensor sits beneath the sample, and on its cover glass — facing the pixels — the team fabricates a disorder-engineered coded surface: a random layer of sub-micron scatterers that modulates both intensity and phase.

The sample floats 0.2 to 2 mm above the surface and never touches it, so the code tolerates contamination and extended operation, including inside a cell incubator. The team calibrated the system in advance with a static blood smear.

During acquisition, voice-coil actuators translate the coded sensor beneath the sample, and a static fiducial marker tracks position with sub-pixel accuracy. Because the code scrambles fine spatial frequencies into diffraction patterns the pixels can record, every frame carries detail finer than the pixel pitch, which ptychographic phase retrieval recovers. With no lens, resolution and field of view decouple: the field of view is simply the sensor's active area.

How does it film moving objects?

Moving the sensor solves half the problem. A dynamic object changes between frames, which breaks classic ptychographic algorithms built for static scenes. The team's observation: natural dynamic scenes are highly correlated — structures persist across frames and change along smooth paths in time.

Instead of storing each frame, a neural field holds one shared spatial map plus roughly 10 temporal keyframes. A small neural network combines the two to predict the light field, including phase, at any point and moment. The fit requires no training data; the model learns directly from each experiment's measurements by simulating how the predicted light passes through the coded surface and adjusting until the simulation matches every recorded frame.

Every frame then does two jobs: pushing spatial resolution toward gigapixel scale and sampling the dynamics at video rate. A 50-second recording at 30 fps — about 1,500 raw frames — compresses from gigabytes into a megabyte-scale model. In this scheme, SBP grows with how much correlation the reconstruction extracts from the sample dynamics, not with measurement or sensor count.

Where does the measured performance land?

The recovered phase tracked the melting front of a snowflake and the thickness of its meltwater film while the finest dendrites stayed resolved. The team also filmed crystals growing from solution, from first dendrites to a polycrystalline film spanning the whole sensor.

For label-free biology, quantitative phase imaging avoids fluorescence phototoxicity. Placed in a standard incubator, the device imaged bone marrow mesenchymal stem cells during a wound-healing assay; the height maps resolved tunneling nanotubes — membrane bridges that usually require fluorescent labels — and followed cell division, cargo transport along nanotubes, and necrosis after hypotonic shock. Cells that drift out of focus can be refocused digitally after acquisition, because the reconstruction recovers the full complex field leaving the sample.

Two further demonstrations target clinical and pharmaceutical use:

• Under gentamicin, some E. coli cells in a microfluidic channel grew into long filaments instead of dividing; phase tracked culture dry mass with picogram sensitivity — a readout for rapid antibiotic susceptibility testing.
• On a transdermal microneedle patch, unwrapped phase confirmed the 700-μm design height across the patch, then followed each needle as it dissolved.

Does the method hold beyond visible light?

The framework does not require the object to be the part that changes. In extreme-ultraviolet ptychography, source instabilities make the probe beam fluctuate between exposures. On 250 diffraction patterns from a tabletop EUV source (Eschen et al., 2023) with only 24% probe overlap — well below the usual 50% — the method recovered the sample and the frame-by-frame probe evolution in 40 iterations, while conventional orthogonal probe relaxation still showed artifacts after 1,500. The object stayed recognizable at 1.2% overlap.

That robustness matters most for x-rays and electrons, where radiation damage progressively alters the sample during acquisition and forces a dose-versus-resolution trade. Modeling the evolution at camera frame rate could allow full-dose imaging with the structural change accounted for.

The work was partially supported by the U.S. Department of Energy (SC0025582) and the National Institutes of Health (R01-EB034744).

The recipe generalizes: a code that spreads information across time, and a model that recovers it from correlations. The authors point to event-based sensors, multiwavelength imaging, fiber-bundle endoscopes, and long-range imaging through turbulence. The open question for test-instrument buyers: whether correlation-driven SBP, which depends on scene dynamics the operator does not control, can be specified with the guaranteed accuracy classes that metrology labs demand.