Retinal and Semantic Segmentation of Diabetic Retinopathy Images Using MobileNetV3

The eye is affected by diabetic retinopathy (DR), a condition caused by diabetes. DR may initially show no symptoms or cause minor vision problems. However, it may lead to blindness if not diagnosed and treated early. The goal of this research is to segment and classify DR into five stages: no DR, mild, moderate, severe, and proliferative retinopathy. In this work, a deep learning-based convolutional neural network (CNN), namely, MobileNetV3 is employed on a set of 5590 images to identify the stages and patterns of DR. For this, accurate retinal vascular analysis is necessary. This can be accomplished by retinal segmentation to provide a precise result. Retinal segmentation is the process of automatically identifying blood vessel borders. It is designed to handle heavy-duty use cases as well as low-resource use cases. Here, retinal segmentation is carried out by U-net architecture. Due to the region merging process, characteristics loss in the segmentation are preserved and passed on to the image classifier, which has an accuracy rate of up to 97%. MobileNetV3-Large and MobileNetV3-Small are both deep neural networks designed for heavy and low-resource uses, respectively. A model with variable parameters is tweaked and used to identify objects and perform semantic segmentation tasks. Implementation results show that MobileNetV3-Large is 3.2% more accurate than MobileNetV2, while the latency has been reduced by 20% compared to MobileNetV2. MobileNetV3-Small is 6.6% more accurate than the MobileNetV2 model with equal latency.