Publication

Abstract : Diabetic retinopathy (DR) grading is significant and impactful in the detection and prevention of impaired vision over time. In this project, we have deployed a hybrid fusion methodology to represent a combination of deep features of pretrained Efficientnet variant models and handcrafted colour & texture descriptors for DR classification based on severity. It was tested on the APTOS 2019 retinal fundus image dataset which does have severe class imbalance. The results of this approach suggest that Random Forest (RF) classifier has best test accuracy of 73.41% and Cohen kappa of 0.8411. It outperforms the other classifier combinations by more than 2%. The use of class-weighting methods helps address class imbalance and improves sensitivity across DR severity levels. These results demonstrate the potential of using different feature representations for DR and future integration with multimodal clinical retinal data.