An efficient model for detecting retinal diseases from fundus images using deep learning

Detection of retinal abnormalities is an interesting problem that has attracted the attention of several studies. Although many works have been published on this topic, they are still ineffective when applied to real data from medical facilities due to conditions and constraints that lead to low efficiency. To address this problem, this paper proposes an effective model to detect signs of retinal diseases, including glaucoma, cataracts, and diabetic retinopathy. The proposed model EfficientNet-TL combines transfer learning techniques with the EfficientNet network architecture to improve accuracy. In addition, EfficientNet-TL also applies the histogram of oriented gradients (HOG) technique to extract information about the structure and shape of abnormal regions on fundus images and the LBP technique to extract image texture features and recognize small changes on the retinal surface. Experimental results on the image dataset showed that the proposed model EfficientNet-TL achieved high efficiency with an accuracy of up to 93%, which is superior to the previous models.