Glaucoma is a chronic eye disease which causes progressive damage to the optic nerve and leads to blindness if not treated. This work demonstrates that an automated AI system can accurately and remotely diagnose glaucoma from fundus images of the retina of the eye. The project is comprised of two main parts. In the first part, raw digital fundus images undergo pre-processing to prepare them for analysis. In the second part, the processed images are analyzed by the convolutional neural network and the system outputs a diagnosis and a confidence level. The system includes an intuitive interface that allows users to easily upload images and understand the output.
Glaucoma is the second leading cause of blindness in the world, with 70 million people afflicted worldwide and around 80,000 patients in Israel alone. It is a chronic eye disease which causes progressive damage to the optic nerve and leads to blindness if not treated. Glaucoma can be diagnosed from digital fundus images of the eye through manual analysis of several parameters, including the cup to disk ratio.
This work shows that glaucoma can be accurately diagnosed from fundus images by a convolutional neural network-based classifier. The classifier is based on the AlexNet architecture, developed in 2012. The project is comprised of two main parts.
In the first part, raw digital fundus images undergo pre-processing to prepare them for analysis. The images are segmented, the relevant region containing the cup and the disk is extracted and saved with a uniform aspect ratio. The pre-processing is necessary because the images that the neural network analyzes must be as uniform as possible to achieve optimal diagnosis accuracy.
In the second part, the processed images are analyzed by the convolutional neural network and the system outputs a diagnosis and a confidence level. In this project, it is also demonstrated that the system can be implemented in software only or deployed to an FPGA. The deployment is performed using the Vitis-AI suite of tools provided by Xilinx, and the target system is the Xilinx ZCU-104. The ZCU-104 contains enough memory and computational power to make it suitable for implementation of the CNN. FPGA deployment has the advantage of accelerating the system and increasing portability.
