CNN based screening system to discriminate healthy retina

The detection of eye disease starting from the observation of the retina is crucial for preventing partial or permanent blindness in patients. In recent years, the number of studies for the automatic diagnosis of such diseases related to the retina has increased. Consequently, the scope of this thesis is to generate a system that, given an image of a retina perceived from an ophthalmoscope, is able to predict if a patient has a healthy eye or if there is a disease. Moreover, the output will also contain a heatmap, illustrating the area of the image that was considered by the system to make the prediction. For this purpose, the first step is to gather the dataset of retinal images and apply some pre-processing steps, in order to highlight the elements that are related to a certain disease. Then there are two main different approaches used for this study: the first one is to construct a CNN from scratch with a simple architecture for a binary classification task, and the second one uses the technique of transfer learning. For the second approach the starting architectures used and compared together are VGG16, InceptionResnetV2, EfficientNetV2 and DenseNet121. At the end all the techniques will be compared based on qualitative and quantitative results to show which one is the best solution for this specific problem.