Dry eye disease (DED)
is one of the most common ophthalmic disorders
. Inflammation of the ocular surface is controlled by corneal antigen-presenting cells called dendritic cells (DCs)
, which induce T-cell activation, and play a critical role in the pathogenesis of dry eye disease. The density of corneal DC is correlated with both symptoms and clinical signs in dry eye, making it an attractive non-invasive and responsive surrogate biomarker to assess the severity of corneal inflammation. To evaluate the effect of inflammation on the cornea
structures and function, coronal images at a cellular level are acquired in a non-invasive manner by in vivo confocal microscopy (IVCM)
, to visualize the density of immune and inflammatory cells. Manual classification immune cells in IVCM images is highly time consuming and prone to subjective decisions and levels of expertise: it therefore becomes a bottleneck in the detection and assessment of DED prognosis. However, an automated convolutional neural networks (CNN) for the detection and quantification of dendritic cells in IVCM images can be used as an objective evaluation of ocular surface inflammation to improve both diagnostic and treatment accuracy.
For this classification and segmentation of dendritic cells, RSIP vision in collaboration with physicians from Tufts Medical Center
have trained a U-Net based CNN to classify corneal layers, segment and quantify DCs density in four tissue types acquired by IVCM images: epithelium, subbasal nerve plexus (SBP), stroma, and endothelium. U-Net is a fully CNN architecture, which takes as an input arbitrary size images (A in the figure below) and produces corresponding-sized output with segmented object in an efficient procedure, comprised of successive steps of down sampling and up-scaling (B in the figure below). We have trained the U-Net to distinguish DC on images containing the four cell types. To generate the training set, dendritic cells were first identified in IVCM images, using image processing techniques, as bright individual dendritiform structures with cell bodies and were tested against manually segmented cells by trained experts.
Classification and segmentation of dendritic cells – Results
We have obtained 97% accuracy using our U-Net and demonstrated that deep learning can be utilized in the analyses of laser IVCM images, allowing a standardized, objective, and rapid evaluation of ocular surface inflammation. In the attached animated movie, we show an example of IVCM images, an overlay of the U-Net segmentation results (false negative – red, true positive – green, false positive – blue) and ground truth segmentation (white), which are in excellent agreement. Using the automated segmentation, a rapid evaluation of the spatial distribution and density of Dendritic Cells can be given with high accuracy.