Assessment of XAI Tools in Interpretation of Diabetic Retinopathy Grading

Diabetic Retinopathy is an eye disorder caused due to access amounts of sugar in the blood which creates blockages in the vessels that supply blood to the eyes. This damage caused to the retinal blood vessels is a result of diabetes and can lead to vision loss. If this disorder is untreated, it can cause partial vision loss or even total blindness within a few years. Hence, early detection and treatment of Diabetic Retinopathy is very important. In our project, firstly a pre-trained deep learning model (ResNet-50) was used to classify a given Diabetic Retinopathy fundus image into one of the five classes (No DR, Mild DR, Moderate DR, Severe DR, Proliferative DR). The output produced was a black box output. In the medical field it is very important that proper explanations are provided as to why an AI model has given a certain output. Hence, we built two Explainable AI models that will help doctors as well as patients identify the features that have led to the classification of a particular DR fundus image in a given category with the of Heatmap generation. The two Explainable AI models built were Grad-Cam and Grad-Cam++. The Grad-Cam model provided a coarse localization map that highlighted the major areas in the image for estimating the target features. Whereas the Grad-Cam++ model produced a superior localization of the objects. After the two models were implemented, the comparison of the results obtained from the two models showed that the localization of the features in Grad- Cam++ were far better than Grad-Cam. It was also observed that Grad- cam++ was able to identify multiple objects of a single class with better accuracy whereas Grad-Cam was not able to do the same. Hence, concluding that Grad-Cam++ performed better than Grad-Cam.