AssessCataract : Ensemble Learning Approach for Cataract Severity Assessment using Digital Images

Abstract

"Cataracts, the primary cause of blindness globally, occur when protein structure of the lens becomes clumped together to form a white cloud that leads to gradual vision loss. While age is the primary risk factor, external factors such as smoking and drinking will increase the risk of cataracts. Early stages of cataracts can be maintained with the use of special glasses or contacts, but surgical replacement of lenses is required for critical stages. Timely identification of cataracts can improve and enhance patient outcomes and provide more personalized treatment strategies. Many studies focus on image processing and deep learning techniques to detect and grade cataracts which often use fundus images for classification but the devices to capture these images are expensive and not available in most underserved areas. The research conducted using digital images also has several limitations and potential research gaps. Our proposed approach contains a CNN model for binary classification and an Ensemble model for severity assessment of cataract. This CNN model is composed with four convolutional layers followed by pooling layers, single flatten layer and finally, two dense layers and several pretrained models were trained and tested and the highly performing models, VGG19, InceptionV3 and MobileNet were chosen as the base models for the Ensemble model. The cataract detection model was trained and tested, achieving a testing accuracy of 95%. The severity assessment model, an ensemble multiclass classification model, achieved an accuracy of 78% after training and testing. These accuracies represent significant progress addressing the challenges and limitations of cataract detection and severity grading systems."