Berkeley details AI-assisted imaging that reveals skin cell differences without dyes
The proof of concept pairs stable optical hardware with AI processing, measuring phase variations approximately 10 times smaller than a previously reported benchmark.
UC Berkeley researchers demonstrated a label-free optical imaging approach that distinguished normal and cancerous skin cell lines by their responses to light across wavelengths. The proof of concept pairs a more stable interferometer with AI processing and reports phase-measurement sensitivity about 10 times smaller than a previously reported benchmark. The result points to a possible route to sensitive cell analysis without dyes, but disease detection remains an application to develop and validate; the team’s next goal is a more compact instrument.
01
The method, General Polarization Common-Path Interferometry (GPCPI), uses closely aligned reference and sample beams to reduce noise from shared environmental disturbances.
02
Neural autoencoders flag irregularities in phase spectra, while ConvNeXt V2 detects optical changes without reconstructing a full spectrum for every measurement.
03
The cell-imaging demonstration compared normal CCD-32Sk cells with cancerous COLO-829 cells; it does not establish a validated diagnostic.
Normal and cancerous skin cell lines produced distinct optical fingerprints without chemical labels or fluorescent dyes in a UC Berkeley proof of concept described November 20. The technique combines closely aligned light beams with AI analysis to measure subtle changes in light that reveal properties of the samples.
Berkeley’s November 20 description says the findings were reported in Optica; the authors’ preprint was submitted January 2, 2026, and revised July 20.
Steadier light, then smarter processing
The method is called General Polarization Common-Path Interferometry, or GPCPI. Interferometry combines light beams to reveal changes in phase—the timing of a light wave’s oscillation. Those changes carry information about a sample’s physical and optical properties, beyond what brightness and color alone can show.
Sensitivity creates a problem: vibrations and other environmental disturbances can introduce noise. GPCPI splits light into two closely aligned beams. One acts as a reference; the other probes the sample. Their nearly identical paths expose them to similar disturbances, making measurements more stable.
The design also eases conventional common-path systems’ restrictions on samples’ polarization properties, broadening the materials that can be examined. AI supplies a separate layer of signal processing.
AI has two jobs
Correct the measurement: AI helps identify and correct irregularities in phase spectra—measurements of phase across different wavelengths. Neural autoencoders detect anomalies in those spectra.
Read the interference pattern: a deep-learning model detects subtle optical changes without reconstructing the full phase spectrum for every measurement. The authors identify their adopted model as ConvNeXt V2 and describe real-time phase tracking with reduced noise.
Phase measurement sensitivity
Approximately 10 times smallerMeasured phase variations
The researchers report this improvement over a previously reported benchmark using the combined optical architecture and AI analysis.
From gold structures to skin cells
Researchers first used nanoscale gold structures to sense changes in the optical properties of surrounding liquids. They then applied the platform to hyperspectral cell imaging: measuring how cells’ responses to light change across different wavelengths.
The normal skin cell line was CCD-32Sk; the cancerous line was COLO-829. Lead author Kamyar Behrouzi says avoiding labels and fluorescent dyes matters because they can alter biology or damage living samples. Cell classification and disease detection remain potential applications requiring further development and validation.
The team’s next engineering goal is a more compact, practical optical platform. Its stated ambition is to make sensitive, AI-assisted measurements accessible across scientific, biomedical and industrial work.
Sources
arxiv.orgAI-Assisted Hyperspectral Interferometry and Single-Cell Dispersion Imaging
Reader comments
Newest comments first. Replies stay oldest first.