Transfer learning based robust automatic detection system for diabetic retinopathy grading

Abstract

Diabetic retinopathy (DR) can be categorized on the basis of prolonged complication in the retinal blood vessels which may lead to severe blindness. Early stage prediction and diagnosis of DR requires regular eye examination to reduce the complications causing vision loss. Indicative significance of DR forecast and evaluation to help the ophthalmologists in standard screening has prompted the improvement of computerized DR recognition frameworks. This work focuses on automatic DR disease identification and its grading by the means of transfer learning approach using dynamic investigation. Our proposed approach utilizes deep neural network for feature extraction from fundus images and these features are further ensembled with supervised machine learning technique for DR grading. An optimized classification is achieved by applying an ensemble of convolution neural networks (CNNs) with statistical feature selection module and SVM classifier. The learning of classifier is achieved by the feature information transferred from CNN model to the SVM classifier, which results in remarkable performance of the learned models. Statistically optimized feature set utilized for transfer learning technique yields in the classification accuracy of 90.51% with proposed Prominent Feature-based Transfer Learning (PFTL) method employing Inception V3 model. The cost analysis of the proposed model provides a minimum cross-entropy loss of 0.295 consuming the time of 38 min 53 s, thus, maintaining a trade-off. The generalization ability of the proposed model is established by the performance assessment using latest IDRiD dataset that yields accuracy of 90.01% for Inception V3 network providing uniform outcomes for all the evaluation parameters. The diagnosis ability of the proposed transfer learning-based model is justified by comparing the proposed methods with the state-of-the-art methods. The optimized PFTL model (CNN ? Statistical Analysis ? SVM) outperforms other classification algorithms and provides the maximum accuracy improvement of 16.01% over the state-of-the-art techniques.