Automatic Detection And Segmentation Of Retinal Manifestations Due To Diabetic Retinopathy

In the world, diabetic retinopathy is a major factor of preventable blindness. Early treatment plans might lessen the burden of diabetic retinopathy-related blindness, especially in persons in developing nations who are of working age and the elderly. The early signs of diabetic retinopathy include retinal anomalies, and it is crucial for early diagnosis to accurately identify aberrant regions using segmentation. It takes a lot of time and effort to manually identify each and every aberration in diabetic retinopathy. In this study, we suggest a technique for automatically segmenting retinal anomalies into multiple classes using semantic multiclass segmentation. Exudates, microaneurysms, and haemorrhage are among the anomalies taken into account. The 81 colour fundus images that make up the IDRID dataset, mask were carefully created under the direction of an ophthalmologist. The dataset is then fed into the U Net architecture using Efficientnet as the encoder for the Unet. With a loss of 0.0969 and an IoU score of 88.26%, this work achieved accuracy and precision of 99.65% and 99.33%, respectively. This work provides a truly automated technique for detecting diabetic retinopathy that can also rate the quality of the retinal image. Automated diagnosis is considered to provide advantage in processing a lot of images efficiently, reliably, systematically, and economically.