Authors - Ahmed Noorim, Raina Nusrat Jahan, Md. Sabbir Al Ahsan, Sourav Adhikary, Md. Jamil Uddin Abstract - Early detection of Diabetic Retinopathy (DR) remains essential due to its status as a leading cause of vision loss along with severe complications. Ever increasing worldwide diabetes situation makes it necessary to develop an automated diagnosis system for detecting DR at an early stage precisely. The research proposes a weighted average ensemble deep learning architecture to perform two stages of DR diagnosis along with severity classification from fundus images. The first stage identifies cases of the presence of DR by applying a binary classifier followed by a multiclass classifier in the second stage to evaluate severity levels. This model is trained and evaluated on a merged dataset which amalgamates APTOS 2019, MESSIDOR 2, and IDRiD with three different preprocessing to boost its generalized application capabilities. DenseNet121, EfficientNetB1 and Xception models complement each other for optimal feature extraction and classification task thus utilized in the development of ensemble model. Outperforming several state-of-the-art models, 97% recall with 92% accuracy was achieved in determining the existence of DR, while severity classification reaches 93% accuracy and 94% recall. The research shows promising assistance for ophthalmologists in becoming an essential diagnostic screening instrument for the early detection of DR in resource limited areas.
Open Zoom
