A lesion based diabetic retinopathy detection through hybrid deep learning model

Diabetic Retinopathy (DR) is a condition that is highly prevalent in diabetes and a leading cause of blindness and impaired sight in the rest of the world. The timely treatment is only possible by early and proper diagnosis of DR in order to avoid a total loss of vision. The retinal lesions like microaneurysms, hemorrhages, exudates, and cotton wool spots are significant pointers to be used in identifying the extent to which the disease has progressed.In this paper, the author suggests a lesion-based diabetic retinopathy detection system built on a hybrid deep learning structure comprising of VGG16 and Long Short-Term Memory (LSTM) networks. Within the framework of the given method, VGG16 is used as a deep convolutional extractor of features used as discriminants of spatial features of the retina fundus images. These characteristics record significant trends that are linked with retina lesions. The deep features are extracted and then fed to an LSTM network to work out contextual and sequential connections between lesion patterns, making it possible to classify diabetic retinopathy stages more accurately.
It is proposed that the hybrid architecture will combine spatial feature extraction with sequential feature learning, which will enable the system to efficiently discriminate among various phases of diabetic retinopathy like normal, mild, moderate, and severe. The experimental findings prove that the proposed VGG16LSTM model can be effectively used to obtain a higher accuracy, sensitivity and specificity than traditional CNN-based models. The findings are that lesion-based feature extraction with sequential modeling is more reliable in the automated diagnosis of diabetic retinopathy. The suggested system will be able to help ophthalmologists identify diabetic retinopathy at an early stage and evaluate its severity, thus helping them to provide clinical treatment in time and minimize the risk of patient vision loss.