A unified technique for entropy enhancement based diabetic retinopathy detection using hybrid neural network. (June 2022)

• Record Type: Journal Article
• Authors: Fatima, ; Imran, Muhammad; Ullah, Anayat; Arif, Muhammad; Noor, Rida
• Is Part Of: Computers in biology and medicine. Volume 145(2022)
• Publication Date: 2022-06
• Subjects: Diabetic retinopathy -- Neural network -- Discrete wavelet transform -- Histogram -- Classification -- Medical data; 41A05 -- 41A10 -- 65D05 -- 65D17; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2022.105424
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: In this paper, a unified technique for entropy enhancement-based diabetic retinopathy detection using a hybrid neural network is proposed for diagnosing diabetic retinopathy. Medical images play crucial roles in the diagnosis, but two images representing two different stages of a disease look alike. It, consequently, make the process of diagnosis extraneous and error-prone. Therefore, in this paper, a technique is proposed to address these issues. Firstly, a novel entropy enhancement technique is devised exploiting the discrete wavelet transforms to improve the visibility of the medical images by making the subtle features more prominent. Later, we designed a computationally efficient hybrid neural network that efficiently classifies diabetic retinopathy images. To examine the effectiveness of our technique, we have chosen three datasets: Ultra-Wide Filed (UWF) dataset, Asia Pacific Tele Ophthalmology Society (APTOS) dataset, and MESSIDOR-2 dataset. In the end, we performed extensive experiments to validate the performance of our technique. In addition, the comparison of the proposed scheme – in terms of accuracy, specificity, sensitivity, precision and recall curve, and area under the curve – with some of the best contemporary schemes shows the significant improvement of our techniques in terms of diabetic retinopathy classification. Highlights: In this paper, we have: Proposed an entropy-based enhancement technique to improve the visibility of medical images better and to make the subtle features more conspicuous for better classification results. Proposed a novel hybrid neural network architecture that is computationally efficient and gives good results when trained on smaller or imbalanced datasets. Tested the performance of the proposed methodology using three datasets: UWF, APTOS, and MESSIDOR-2. Compared the proposed technique with state-of-the-art techniques and obtained better results.