Ensemble classification based optimized transfer learning feature method for early stage diagnosis of diabetic retinopathy

Early diagnosis of the diabetic retinopathy is important in preventing vision loss. The intense workload of the experts and the long time analysis of the fundus images necessitate a fully automatic early diagnosis system. For this purpose, a new strategy, which applies preprocessed images to Mobilenet transfer learning and motivated by the ensemble learning trained with Minimum Redundancy Maximum Relevance feature selection is proposed. Unlike other studies, early diagnosis of diabetic retinopathy is more important. For this reason, this early detection was chosen as the focus of the study. The performance of the proposed method was validated by the Messidor, Aptos 2019 and DDR datasets including the first stage of diabetic retinopathy images. The labeled data was arranged by the synthetic minority over sampling technique because the image features obtained by transfer learning have an unbalanced distribution. The ensemble classifier trained with the features selected by Minimum Redundancy Maximum Relevance for classification. Experimental results proved that the proposed method significantly increases the detection accuracy for the first stage of diabetic retinopathy for early diagnosis. The early diagnosis achieved accuracies of 98.6%, 88.95%, 97.46% and false positives of 3, 30, 30 on Messidor, Aptos 2019 and DDR datasets, respectively. Also multiclass classification of all stages of diabetic retinopathy were gathered as 96.06%, 89.86%, 82.74% on Messidor, Aptos 2019 and DDR datasets, respectively. This result proved that the proposed method is capable of assisting experts working in the field of ophthalmology to automatically analyze fundus images in early diagnosis. © 2023, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.