An artificial intelligence-driven workflow to detect the macula and quantify retinal ganglion cells in nonhuman primates.

The retinal ganglion cell layer integrates and transmits stimuli from photoreceptors to the central nervous system. Retinal ganglion cell loss is a hallmark of glaucoma and other retinopathies and neuropathies. Therefore, changes in this cell population define disease severity and therapeutic efficacy in patients and animal models. Artificial intelligence has demonstrated utility in automated quantification of whole-slide images, including hematoxylin and eosin-stained tissue sections. Hematoxylin and eosin is the most commonly used microscopy stain, easily implemented, highly reproducible, cost-effective, and able to delineate cellular details; however, manually counting cells is labor-intensive and subject to interobserver variability. We have applied artificial intelligence to automate detection of the macula and retinal ganglion cells at the level of the fovea in 4-µm-thick cross sections of African green monkey (Chlorocebus sabaeus) eyes. The workflow finds the parafoveal and perifoveal macula versus the peripheral retina and quantifies the number of retinal ganglion cells in the specific region of interest. Results were validated by comparing pathologist-guided manual annotations to the artificial intelligence outputs using accuracy and intersection over union (IoU). The macula-segmentation application achieved 97.1% accuracy and an IoU of 0.89, whereas the retinal ganglion cell counting application showed strong agreement with pathologist counts (R2 = .9986). This study provides a proof of concept for automated detection and quantitation of macula and retinal ganglion cells in hematoxylin and eosin-stained whole-slide images.