Diabetic retinopathy detection and diagnosis by means of robust and explainable convolutional neural networks

The diabetic retinopathy is a disease affecting the retina and it is currently manually diagnosed by specialists. In order to help the clinician in this time-consuming task, we propose a method aimed at automatically identify the diabetic retinopathy presence from ocular angiography by exploiting convolutional neural networks. In particular, two models are proposed: the first one is aimed to discriminate between healthy eyes and eyes with retinopathy, while the second one is designed to distinguish between non-proliferative retinopathy and weakly and severely proliferative retinopathy. The results we obtained, i.e., an accuracy of 0.98 for the first model and an accuracy of 0.91 relative to the second model, demonstrate that the proposed models can effectively aid the clinician in diagnosis. Moreover, the proposed method is aimed to localize the disease in the angiography, providing a kind of explainability behind the model diagnosis, by taking into account two different class activation mapping algorithms showing on the images the areas symptomatic of the disease, in order to increase model trustworthiness from doctors and patients. We also introduce a similarity index aimed to evaluate the model robustness by quantifying how much the heatmaps generated by the class activation mapping algorithms of the same model differ from each other.