Robust Exudate Segmentation in Retinal Fundus Images using the combination of Efficient-Net and U-Net Model

Diabetic retinopathy (DR) is an eye disease. Lesions like exudates play a major role while grading DR. Hence, it’s important for accurate segmentation of exudate, as it plays a significant role in detecting and segmenting lesions. The exudates are smaller in shape and size, which are often irregular, and the contrast is also not high. Therefore, it is challenging to identify the exudate’s actual location to get exudates in fundus images. Here in The paper, we present a novel deep neural network architecture that focuses a combination of EfficientNet and U-net architecture to accurately segmenting the exudate present in the fundus image. The proposed framework combines a U-shaped structure with residual connections to capture the subtle features of exudates. Here, in this paper, we proposed model has been utilized using Karas and Tensor-Flow tools, and it provides accurate results in identifying exudates, even small in size. This model has been trained and evaluated on the Indian Diabetic Retinopathy Image Dataset (IDRID). The experimental results show a novel precision (pre) value of 96.79%, Area Under Curve(AUC) value of 90.57%, a dice coefficient value of 96.87%, and an IOU value of 95.63%, This method demonstrates a significant improvement, outperforming the other states-of-art approaches in terms of accuracy and performance or efficiency.