摘要：Cases of diabetes and accompanying diabetic retinopathy (DR) have also been developing at an accelerating rate in recent times. Early identification of DR is a significant concern as that might cause irreversible blindness in the advanced stages. In the past couple of decades, numerous alternative methodologies have been utilized in DR diagnosis. Literature demonstrates that perhaps the deep neural networks were the most recommended technique for DR detection. Among such techniques, Convolutional Neural Network (CNN) architectures are the most utilized ones in the area of medical imaging classification. Constructing Classifier model is a difficult and time-consuming process. Additionally, training a vast number of parameters is also a challenging effort. Due to this rationale, so rather than training CNNs from beginning, employing pre-trained architectures has been recommended in latest days as transfer learning strategy. This work employed transfer learning technique on pre-trained ResNet50 and built a customized block of CNN stacked on top of ResNet50 for generating the hybrid model. It has been assessed the effectiveness of the proposed scheme using APTOS 2019 blindness detection dataset and obtained a Cohen Kappa score of 96.7 % , 89.3 % and Accuracy of 95.8 % , 83.4 % on train and test data, correspondingly. This model performed much better than some of the other reported findings to yet.
关键词：This model performed much better than some of the other reported findings to yet
