Retinal layer thickness prediction from smartphone fundus imaging for early glaucoma detection

Glaucoma progression leads to the loss of ganglion cells, which results in the thinning of retinal layers. Retinal layer thickness is a biomarker for monitoring glaucoma progression. Glaucoma is detected using the thickness of OCT images based on the RNFL layer and fundus images based on optic nerve damage. Continuous glaucoma monitoring requires both screening methods for effective treatment. However, OCT devices are high-cost, non-portable, require specialized operators, and are susceptible to artifacts and media opacities, which reduce their suitability for routine, large-scale screening, particularly in resource-limited settings. To address this clinical accessibility gap, an MPIML framework is proposed. The framework uses the Glauco Twin dataset, which consists of both fundus and OCT image data features. From the dataset, RNFL layer thickness is predicted from fundus image data features. This method of prediction avoids frequent OCT screening. In this proposed study, a smartphone-based 2D fundus image is used for measuring RNFL thickness. As the current cohort comprises 23 patients, this work is presented as a pilot feasibility study intended to demonstrate technical and clinical feasibility rather than population-level clinical validity. In this study, a multimodal physics-informed machine learning (MPIML) framework is proposed for the measurement of retinal layer thickness. The framework uses (I) the Glauco Twin dataset, which consists of segmented OCT and smartphone-based fundus image features such as structural, vascular, and biomechanical features for the same glaucoma patient's left and right eye. The MPIML framework consists of Segmentation-Modified TransUNet, which segments OCT images with 2-layer boundary segmentation of RNFL and ILM. A Modified SegFormer algorithm segments the optic disc, optic cup, blood vessels, and macula region for estimating indirect RNFL measurements. Finally, the MPIML model is used to classify glaucoma progression as normal, suspect, and glaucoma. The MPIML framework achieves a classification accuracy of about 94.3%, an AUC of 0.95, and RNFL thickness prediction is done for 3.1–3.4 µm length. In this pilot feasibility study, smartphone-based fundus image data with the MPIML model effectively predicts RNFL thickness and improves the detection of early-stage glaucoma progression, motivating a larger, multicenter validation study.