Deep Learning Approach For Detection Of Retinal Abnormalities Based On Color Fundus Images
Batuhan Bulut, Volkan Kalın, Burcu Bektas Gunes, Rim Khazhin
Abstract
In cases where people cannot access regular controls, treatment and care, delaying the diagnosis and treatment of eye diseases such as glaucoma, cataracts, diabetic retinopathy or leaving them to deteriorate unconsciously, may make daily life difficult and even cause blindness. Therefore, automatic examination of fundus photographs is important in terms of providing early diagnosis with fast, objective and consistent image evaluation and helping the application of large-scale scanning programs. This research uses Xception model with transfer learning method to classify images obtained from Akdeniz University Hospital Eye Diseases Department. During the analysis, the Xception model containing 50 different parameter combinations was trained by scanning the appropriate hyper-parameter space for the model. Comparisons were made for the top 9 models with the highest performance. The 4th model reached the highest accuracy rate with 91.39% for the training set, and as for the validation set, the 0th model showed 82.5% of accuracy. In addition, in order to test the performance of the model with an independent data set, open access fundus images were used for test analysis and binary classification AUC (Area Under Curve) values were calculated for 21 different diseases.