Abstract:
The vision depends greatly on the retina, unfortunately, it may be exposed to many diseases that lead to poor vision or
blindness. This research aims to diagnose retinal diseases through OCT images, focusing on Drusen, diabetic macular edema (DME),
and choroidal neovascularization (CNV). A new ensemble model is proposed that proposes new methods and combines them with soft
and hard voting methods, it is based on three sub-models (Custom-model, Xception, and MobileNet). Because we noticed that some
sub-models are better than others at classifying a particular category, each sub-model was assigned to the category it classifies best.
We also used a way to correct final misclassification through a list of negative predictions created to contain categories to which the
sub-model is somewhat certain that an image does not belong. The proposed ensemble model achieved a state-of-the-art accuracy of
100%, and the Custom model obtained an accuracy of 99.79% on the UCSD-v2 dataset. The Duke dataset was also employed to verify
the performance efficiency of the model, with the ensemble model also achieving an accuracy of 100%, and the Custom model
recording an accuracy of 99.69%. In the first dataset, the custom model specializes in Drusen and Normal, Xception in DME, and
MobileNet in CNV. While the custom model in AMD, Xception in DME, and MobileNet in Normal in the second dataset. The results
of this research emphasize the effectiveness of ensemble learning techniques in analyzing medical images, especially in diagnosing
retinal diseases.