Analysis of Pre-trained Convolutional Neural Network Models in Diabetic Macular Edema Detection Through Retinal Fundus Images

Diabetic Macular Edema (DME), a serious complication linked to Diabetic Retinopathy (DR), can result in vision loss and potential blindness. DME occurs when fluid leaks from blood vessels in the macula or when the retina thickens. Fluid leakage is presented as Hard Exudates (HE), which appear as yellow or white clusters of varying shapes, sizes, and positions, serving as indicators for diagnosing DME in fundus color images. Early detection of DME can significantly improve treatment options and patient quality of life. In this study, we propose evaluating three pre-trained Convolutional Neural Networks (CNN) models to assess the risk of DME in fundus color images. In this study, transfer learning is employed to leverage the convolutional base of pre-trained models for feature extraction from retinal images. Subsequently, a custom fully connected convolutional layer is added to perform the classification task. The publicly available MESSIDOR dataset was used to train and test the proposed method, which achieved an accuracy of 95% in detecting DME.