Detection of Hard and Soft Exudates from Fundus Images through Deep Learning
Hassan Aslam, Department of Computer Science, Faculty of Telecommunication and Information Engineering, University of Engineering and Technology, Taxila, Pakistan.
Muhammad Javed Iqbal, Department of Computer Science, Faculty of Telecommunication and Information Engineering, University of Engineering and Technology, Taxila, Pakistan.
Corresponding Author:
Hassan Aslam (hassanaslam852@gmail.com)
Abstract:
The presence of exudates on the retina is a sign that diabetic retinopathy has occurred, and it may be possible to diagnose this disease automatically. Diabetic retinopathy (DR) can sometimes be treated and prevented from causing irreversible vision loss if caught and treated properly. The current research was based on the detection of retinal fundus images which are labeled for hard exudates. The Efficient Net-B5 model on the Asia Pacific Tele-Ophthalmology Society (APTOS) and dataset for Messidor-1 datasets performed exceptionally well when compared to other methods. In current research, an accuracy score measures the effectiveness of the proposed model by using training images 704 of messidor-1 dataset and 2929 images of APTOS 2019 blindness detection (Kaggle dataset). The results are evaluated on 176 test images of messidor-1 dataset and 733 images of APTOS 2019 blindness detection (Kaggle dataset). The model was trained and evaluated using MESSIDOR-1 and APTOS-2019 datasets and obtained maximum test accuracy of 73.7% and 83.97% respectively. This shows that to detect exudates, Efficient Net B 5 can be used to analyze fundus images. By comparing the results to other that also use MESSIDOR-1 and APTOS-2019 dataset, the accuracy is increased.
Keywords:
Diabetic Retinopathy; Deep Learning; Fundus Images; Aptos-2019; Early Treatment Diabetic Retinopathy Study (ETDRS); Messidor-1.