Classification of Ocular Diseases on Fundus Images Using Weighted MobileNetV2

The major cause of blindness in children and adolescents is the ocular disease. It is anticipated that there will be 1.76 billion people worldwide who will lose their eyesight by 2050. However, if the eye condition that caused it can be identified and treated quickly, blindness can be avoided. Therefore, in this research, we proposed a weighted MobileNetV2 to classify ocular disease on fundus images. MobileNetV2 network is chosen because it has a lightweight architecture that allows data to be processed faster. This research also proposed a weighted cost function to improve the performance of the network in handling imbalanced dataset problems. This is needed because usually, the medical dataset has a much lower number of data in the abnormal class than in the normal class, which may cause the deep learning algorithm to fail in detecting the abnormal class. Experimental results on a fundus image dataset that consists of 4 classes ("Normal", "Cataract", "Glaucoma", and "Retina Disease") shows that the proposed weighted cost function can improve the performance of the network on an imbalanced dataset with the accuracy of 66%, precision of 61%, recall of 58%, F1-score of 57%, and running time of 353.81 seconds. Moreover, the proposed weight calculation formula also gives the best performance among other weight calculation formulas.