Smartphone-based retinal image analysis: a convolutional neural network approach for automatic cataract detection

Cataract, the clouding of the crystalline lens that focuses the light entering the eye onto the retina, is one of the most serious eye disease leading to blindness. Early detection and treatment can reduce the rate of complications in cataract patients. This is especially relevant in developing countries where access to healthcare is poor and the lack of eye specialist makes this diagnosis really hard. In this context D-EYE emerges as a smartphone-based ophthalmoscope aims to be efficiently used both by ophthalmologists in clinics, for large screening or in rural areas by not medical personnel. The strength of this device are the possibility to automatically perform diagnosis and the capability of recording and transmitting high-definition images and videos of the fundus oculi. In order to extend the possibilities concerning D-EYE, this project focuses on the development of an algorithm able to automatically detect cataract through retinal images. Several recent studies in literature suggest to use convolutional neural network as a possible solution to this task. For this reason the proposed algorithm is based on MATLAB (The Mathworks Inc., Natick, MA, USA) and in particular a custom convolutional neural network has been implemented using "Deep Learning Toolbox". After an iterative process of refining where different strategies were tested to achieve the best performances, finally the CNN obtains promising results. In terms of classification percentages the leading network successfully classifies 95.9% of the fundus images analysed.