Akademik Veri Yönetim Sistemi
7th International Symposium on Innovative Approaches in Smart Technologies, ISAS 2023, İstanbul, Türkiye, 23 - 25 Kasım 2023, (Tam Metin Bildiri)
Undiagnosed vision disorder affects billions of people. Automatic multiple disease detections have the potential to solve this problem. In this study, studies were carried out on the RFMiD dataset, which consists of images that may contain more than one retinal disease and includes most retina diseases. Studies on disease detection have been carried out by applying the CNN model, in which patch-based attention mechanism is used, to the data set. First, the image was divided into patches in order and the attention mechanism was applied to the obtained patches. Thus, the contribution of these two methods to the performance was examined. In order to prevent overfitting, dropout was applied to the patches at the beginning of the model. In most models, even if the input image size increases after a certain stage, the model reaches saturation. Thanks to the patchbased attention mechanism, the images are given as input to the model without reducing the size of the images much. In addition, the patch-based attention mechanism provides global context between patches. As a result, the best result in the literature was obtained for RFMiD test data with the proposed model.