Applying deep learning on medical data is a very challenging and crucial task. Transfer learning can reduce the cost of training to a great extent by using pretrained deep convolution neural networks. Diabetic retinopathy is the major cause of blindness and it is increasing world-wide at an alarming rate. In this work, we proposes to apply the transfer learning methods for detection of diabetic retinopathy disease and its different stages. We have experimented various deep learning models such as VGG19, ResNet50 and DenseNet201 in order to determine the best classification model for DR detection. The large dataset for diabetic retinopathy consists of imbalance dataset. So this experiment has been performed for both balanced and imbalanced dataset. The results of the models has been analyzed using various metrics such as precision, recall, f1-score and accuracy.
Research Question
RQ: ”How can the transfer learning enhance/improve detection of the different stages in diabetic retinopathy disease ?”
Sub RQ: ”How we can improvise the performance of models (ResNet50, VGG19 and DenseNet201) over imbalanced dataset ?”
Research Objectives
Obj. 1 A Critical Review of Diabetic Retinopathy Detec-tion and Identified Gaps (2014-2019)
Obj. 2 Exploratory Data Analysis to get insight about the feature for Diabetic Retinopathy Detection
Obj. 3 Implementation, Evaluation and Results of ResNet50 (Precision, Recall, F1-Score)
Obj. 4 Implementation, Evaluation and Results of VGG19 (Precision, Recall, F1-Score)
Obj. 5 Implementation, Evaluation and Results of DenseNet201 (Precision, Recall, F1-Score)
Obj. 6 Comparison of Developed Models
** find my project report on - http://norma.ncirl.ie/4316/