A new hyperspectral imaging platform combining metasurface optics with artificial intelligence (AI) may offer a faster, more objective way to diagnose meibomian gland dysfunction (MGD), according to a proof-of-concept study published in PhotoniX.
The system, based on a spectral convolutional neural network (SCNN) chip, achieved diagnostic accuracy exceeding 96% when distinguishing pathological meibomian gland tissue from healthy controls – substantially outperforming conventional RGB image analysis.
MGD is the leading cause of evaporative dry eye disease, yet diagnosis remains heavily dependent on subjective clinical assessments, including gland expressibility, meibum quality, and symptom questionnaires. Early disease can be difficult to detect, and current imaging methods provide limited biochemical information.
The Beijing-based researchers sought to address this gap using hyperspectral imaging, which captures wavelength-specific “spectral fingerprints” linked to tissue composition. Unlike conventional imaging, hyperspectral techniques can reveal subtle biochemical and structural changes invisible to the human eye.
The study analyzed pathological meibomian gland sections from 11 patients undergoing eyelid surgery, including five patients with MGD and six non-MGD controls. Using both commercial hyperspectral systems and a custom metasurface-based SCNN chip, the investigators examined spectral differences across the visible and near-infrared spectrum.
Distinct spectral signatures emerged in MGD tissue: MGD samples demonstrated significantly altered spectral coefficients in the 500–600 nm, 600–700 nm, and 800–900 nm wavelength ranges compared with controls.
The authors suggest these changes may reflect altered hemoglobin dynamics and meibum composition associated with chronic gland inflammation and dysfunction. Notably, spectral abnormalities also correlated with established clinical parameters including tear break-up time (TBUT), meibomian gland expressibility, ocular surface disease index scores, and meibum quality.
To classify MGD, the team trained a neural network using spectral feature maps generated directly on the SCNN chip. The resulting model achieved a mean diagnostic accuracy of 96.22%, compared with 84% for RGB-based convolutional neural network models. A CNN trained using commercial hyperspectral images performed similarly, with 95.88% accuracy, but required far more cumbersome acquisition systems.
The technological advance lies in the SCNN chip itself. Unlike traditional hyperspectral cameras – which are bulky, expensive, and often require lengthy scanning times – the metasurface-based platform performs snapshot spectral acquisition and in-sensor computing within milliseconds.
This miniaturized architecture could have important clinical implications. Current hyperspectral systems are generally impractical for routine ophthalmic use due to acquisition speed and motion sensitivity. By contrast, the SCNN system is designed for real-time operation and may eventually be integrated into slit-lamp platforms for in vivo assessment.
The study also highlights the growing role of spectral imaging in ophthalmology more broadly. While hyperspectral approaches have previously been explored in retinal oxygenation and vascular imaging, this is believed to be the first investigation of meibomian gland spectral characteristics and the first application of optical neural networks to MGD diagnosis.
However, the authors do acknowledge important limitations with their study, including the small sample size and reliance on ex vivo tissue sections rather than live imaging. Further studies involving larger cohorts and in vivo imaging systems will be needed before such a diagnostic tool can become regarded as clinically viable.
Nonetheless, the findings suggest that combining spectral imaging with AI may open a new frontier in ocular surface diagnostics – one capable of moving beyond morphology toward biochemical characterization of disease.
