improved u-net retinal segmentation for disease diagnosis

Retina segmentation is an essential preprocessing step for computer-aided detection and recognition algorithms. as one of the deep learning models, u-net offers the potential to perform classification and segmentation in one step. evaluation of the u-net model for retinal image segmentation was performed on 1200 images from the messidor dataset. we suggest that classification and segmentation tasks be performed in one step. in this case, network training is done based on all the images, and the images that contain disease are entered into the segmentation stage and the segmentation map is extracted. therefore, we used a separate classifier to filter images so that only images containing disease were segmented with u-net. disease-free images remain as output and do not enter the segmentation stage. our approach increased the dice coefficient to 0.8121 and the average accuracy to 72.12. however, our results show that improvements are needed before u-net can achieve a precision comparable to a stand-alone classifier.