Design of a portable retinal imaging module with automatic abnormality detection

摘要

Regular eye examinations are required for the early identification of retinal diseases, but the non-availability of medical resources and time are the main concerns in semi-urban areas with the scenario becoming more critical in rural areas. This paper discusses on the design of a cost-effective universal retinal fundus camera and the development of a novel algorithm for the identification of two prominent vision threatening diseases such as Diabetic Retinopathy (DR) and Age-related Macular Degeneration (AMD) employing the state-of-the-art, Convolutional Neural Networks (CNN) for higher classification accuracy. Furthermore, the proposed system is capable of predicting the disease severity level as well as providing suggestions as to whether an ophthalmic consultation is required for the subject under consideration. Experiments were conducted on images from publicly available databases and hospital repository. The condition Normal-DR-AMD and their severity levels were predicted. The performance scores employed were accuracy (95.83 %), sensitivity (0.93) and specificity (0.97), where the values corresponds to the average scores across different datasets. Furthermore, the developed system predicts the retinopathy stages with an average computational time of about 3.98 s. Both the scores clearly indicate the effectiveness of the proposed algorithm both in terms of correctness of the finding and speed of the generating the medical inference. This model can be extremely useful for rural campaign, where availability of trained medical practitioners is less as well as space for setting up and maintaining an eye care facility are not feasible. (C) 2020 Elsevier Ltd. All rights reserved.