Diabetic Retinopathy Detection: Improving Accuracy Using Multiple Transfer Learning Features from Pre-trained Deep Learning Networks

Diabetic Retinopathy (DR) is a type of complications caused by diabetes. Patients with DR may experience worsening vision, blindness, and eye pain. To effectively address this disorder, DR must be identified and classified according to its severity. Therefore, automated diagnosis of fundus lesions is of great interest for DR early detection. The development of deep learning technology has provided a strong foundation for effective implementation of the automated detection system. In particular, transfer learning techniques have greatly benefited the research community to reduce computation and reuse trained features. In this paper, the outputs from the 'average pooling' and 'fully connected' layers are used as the features to the Support Vector Machine (SVM) classifier with Error Correction Output Code (ECOC). The proposed method outperforms the fine-tuned pre-trained networks in predicting the severity classes with an accuracy of 80.1%. This means that multiple features extracted from the pre-trained networks contribute to a better recognition process.