A Faster RCNN-Based Diabetic Retinopathy Detection Method Using Fused Features From Retina Images

Abstract

Early identification of diabetic retinopathy (DR) is critical as it shows few symptoms at the primary stages due to the nature of its gradual and slow growth. DR must be detected at the early stage to receive appropriate treatment, which can prevent the condition from escalating to severe vision loss problems. The current study proposes an automatic and intelligent system to classify DR or normal condition from retina fundus images (FI). Firstly, the relevant FIs were pre-processed, followed by extracting discriminating features using histograms of oriented gradient (HOG), Shearlet transform, and Region-Based Convolutional Neural Network (RCNN) from FIs and merging them as one fused feature vector. By using the fused features, a machine learning (ML) based faster RCNN classifier was employed to identify the DR condition and DR lesions. An extended experiment was carried out by employing binary classification (normal and DR) from three publicly available datasets. With a testing accuracy of 98.58%, specificity of 97.12%, and sensitivity of 95.72%, this proposed faster RCNN deep learning technique with feature fusion ensured a satisfactory performance in identifying the DR compared to the relevant state-of-the-art works. By using a generalization validation strategy, this fusion-based method achieved a competitive performance with a detection accuracy of 95.75%.

Keywords

• Diabetic retinopathy (DR)
• fundus image
• histograms of oriented gradient (HOG)
• Shear let transform
• faster RCNN
• feature fusion