DIA-VXNET: a framework for automated diabetic eye disease detection using transfer learning with feature fusion network

Diabetic eye disease is a common and significant disorder that affects diabetic individuals, with potentially disastrous eyesight implications if left untreated. Early diagnosis is critical for effective therapy and preventing permanent visual loss. Although retinal fundus pictures are valuable for identifying retinal problems, manual detection can be time-consuming and labor-intensive for ophthalmologists. To solve this problem, a unique automated diagnosis method has been introduced using Deep Learning (DL) to automatically classify four types of diabetic eye disease: normal, cataract, glaucoma, and retina disease. A unique preprocessing method has been developed to enhance the quality of fundus images and improve the accuracy of disease classification. The proposed architecture combines two modern deep learning models, VGG16 and XceptionNET, to achieve great classification accuracy. A transition block is employed to accommodate the different output shapes of VGG16 and XceptionNet, specifically in the last input layer. An additional layer is introduced to retain the main features and merge them with the output of the transition block. Along with the proposed preprocessing technique and architecture, for batch size 128, the system can achieve accuracy, precision, and recall values of (99.76 ± 0.008)%, (98.94 ± 0.016)%, (98.85 ± 0.016)% respectively for diabetic eye disease detection on ‘IDRiD and HRF’ Dataset. The technology solves the challenges ophthalmologists face in manually identifying diabetic eye disease. It reduces the effort needed and enhances the accuracy and speed of diagnosis, offering a practical and effective way to overcome these difficulties.