EfficientNet-based Three Head CNN Model with CLAHE Pre-processing for Accurate Diabetic Retinopathy Detection

Diabetic retinopathy (DR) is an extreme case of diabetes that could lead to lifelong blindness if left untreated in the preliminary stages. Early detection is a challenge and usually requires expert medical inspection of the fundus images. Here we use convolutional neural networks (CNN) to assist in DR detection. Ensembling methods in CNN are known to increase performance and yield better results. A three head CNN model was ensembled to realize classification, regression and ordinal regression tasks as studied in literature. In this paper we modify the three head CNN model to obtain better results for local dataset pre-processed using CLAHE which provides greater feature extraction. This modified model observes an increase in accuracy as compared to the state of the art techniques. Further, dataset collected from Kozhikode District Medical College is used as test set to assess the model for regional differences. The ensembling of heads achieves an accuracy of 96.4% for multistage DR detection. Subsequently binary classification is done using the model to easily predict if the patient has DR or not, achieving an accuracy of 98.36%.