DR-NET: A Stacked Convolutional Classifier Framework for Detection of Diabetic Retinopathy

ResNet-34/DR: A Residual Convolutional Neural Network for the Diagnosis of Diabetic Retinopathy

“…In this paper, the suitability and efficacy of our proposed work for the diagnosis of DR using fundus images were demonstrated. In comparison to other published research [12], [16], [20], [21] summarized in Table 1, our proposed model ResNet-34/DR achieved the highest accuracy 95%.…”

“…Early stages diagnosis contributes to reducing the limitations of human error and allows monitoring the development of the disease, which helps doctors improve medical treatment. In contrast, previous studies [16] and [20] relied solely on DR diagnosis to be normal or abnormal.…”

“…In addition to both studies in [12], [16] are constructed the CNN architecture from scratch. In contrast, our developed model relied on pre-trained TL models, which are more efficient and allow for the comparison of the performance of different networks to determine the best architecture for our problem.…”

“…Since 2015, researchers have relied on CNNs as a powerful tool in the field of computer-aided diagnosis. For example [16], CNN was utilized for the diagnostic of DR in fundus images as one of two classes: normal and abnormal, as such proposed architecture relied on linking three stacks of convolutional filters in parallel, whereby the output is an outcome of global max pooling. The RGB layers of an image were being isolated in order to use the green layer only, as it is the layer that demonstrates the attributes of the disease in the clearest and most distinguishable manner.…”

“…CNNs and TL are the main two methods of automatic DR diagnosis using DL techniques [9], [12], [14]. TL has proven itself as a very effective technique, especially when handling domains of limited data [16]. Instead of completely training a blank network from scratch, a feed forward approach can be used to fix weights in the lower levels that have already been optimized in order to identify the structures that can be generally detected in images, and retain the weights of the upper levels with backpropagation, enabling the model to identify the unique features of a given set of images, such as fundus images, with much less time, training examples, and computational power [14] [19].…”