EDLDR: An Ensemble Deep Learning Technique for Detection and Classification of Diabetic Retinopathy

Diabetic retinopathy (DR) is an ophthalmological disease that causes damage in the blood vessels of the eye. DR causes clotting, lesions or haemorrhage in the light-sensitive region of the retina. Person suffering from DR face loss of vision due to the formation of exudates or lesions in the retina. The detection of DR is critical to the successful treatment of patients suffering from DR. The retinal fundus images may be used for the detection of abnormalities leading to DR. In this paper, an automated ensemble deep learning model is proposed for the detection and classification of DR. The ensembling of a deep learning model enables better predictions and achieves better performance than any single contributing model. Two deep learning models, namely modified DenseNet101 and ResNeXt, are ensembled for the detection of diabetic retinopathy. The ResNeXt model is an improvement over the existing ResNet models. The model includes a shortcut from the previous block to next block, stacking layers and adapting split-transform-merge strategy. The model has a cardinality parameter that specifies the number of transformations. The DenseNet model gives better feature use efficiency as the dense blocks perform concatenation. The ensembling of these two models is performed using normalization over the classes followed by maximum a posteriori over the class outputs to compute the final class label. The experiments are conducted on two datasets APTOS19 and DIARETDB1. The classifications are carried out for both two classes and five classes. The images are pre-processed using CLAHE method for histogram equalization. The dataset has a high-class imbalance and the images of the non-proliferative type are very low, therefore, GAN-based augmentation technique is used for data augmentation. The results obtained from the proposed method are compared with other existing methods. The comparison shows that the proposed method has higher accuracy, precision and recall for both two classes and five classes. The proposed method has an accuracy of 86.08 for five classes and 96.98% for two classes. The precision and recall for two classes are 0.97. For five classes also, the precision and recall are high, i.e., 0.76 and 0.82, respectively.