Cataract disease classification from fundus images with transfer learning based deep learning model on two ocular disease datasets

Abstract

Cataract is one of the most serious eye diseases that can blind if left untreated. Detection of the disease in the early stages rather than in the advanced stages can prevent the patient from being blind. At this point, suspected patients should be constantly checked. Continuous control and follow-up of patients is a tiring and laborious process. For the reasons stated, two different deep learning models are proposed in this article that can be used in the diagnosis and detection of cataracts to assist the work and procedures of ophthalmologists. The proposed deep learning models were run on a fundus dataset with normal and cataract symptoms. The proposed deep learning models provide automatic classification of normal and cataract images. Fine-tuning and layer additions were performed on the upper layer using a pre-trained deep learning model called MobileNet V3 Small. A basic MobileNet V3 Small model has also been created to evaluate the performance of the model, which has been enriched by fine-tuning and adding layers to its upper layers. The difference between the proposed model and the basic model is demonstrated by comparing the classification performances of cataract and normal images with accuracy and complexity matrix measurements. According to the best results obtained in the performance comparisons made by separating the training and test data according to the KFold option, the proposed model gave a more successful result graph of 8.26% than the basic model. Finally, the proposed MobileNet V3 model has also been tested on images composed of two different datasets. On average, the proposed MobileNet V3 model on the combined dataset reached 96.62% accuracy.