PROJECT TITLE :

Explainable Diabetic Retinopathy Detection and Retinal Image Generation

ABSTRACT:

Deep learning has demonstrated successful performance in classifying the label and severity stage of certain diseases; however, the majority of deep learning models offer very few explanations on how to make predictions. We propose to exploit the interpretability of deep learning application in medical diagnosis. Our motivation for doing so comes from Koch's Postulates, which serve as the foundation for evidence-based medicine (EBM) to identify the pathogen. We are able to determine the symptoms that a diabetic retinopathy (DR) detector identifies as evidence to make predictions by isolating neuron activation patterns from a DR detector and visualizing them. To be more specific, we first establish brand new pathological descriptors by employing the activated neurons of the DR detector as an encoding mechanism for information regarding the location and appearance of lesions. Then, in order to visualize the symptom that was encoded in the descriptor, we propose a new network called Patho-GAN. It is designed to generate retinal images that are medically plausible. If we manipulated these descriptors in the right way, we could even control the position, quantity, and categories of the lesions that were generated arbitrarily. We also show that the symptoms directly related to the diagnosis of diabetic retinopathy are present in the synthetic images that we created. The images that were produced by our current method are qualitatively and quantitatively superior to the ones that were produced by the older methods. In addition, in contrast to other methods that can generate an image, but only after a period of several hours, our second level speed has the potential to be an efficient solution for data enhancement.