Project Name: Diabetic Retinopathy Blindness Detection

Project Description: Diabetic Retinopathy is a condition where Blindness may occur to few diabetic patient. The severity lever of DR is classified into 5 classes or stages where the 0 represents no DR and 4 being highest. The idea is to develop a deep learning approach that can detect the level of severity using retina images.

Industry: Medical

Key Contributions: worked with different range of datasets with low resolution 224*224 images. Literature survey with some previous state of the art deep learning approaches, learned a lot from them. Applied various data pre-processing techniques such as image resizing, gaussian blur, circular crops, Data Augmentation etc. Experimented with ResNet-50 model , pretrained with IMAGNET dataset, Finetuned it and tested it with our Retina images. (Achieved around 51% accuracy) Tried to figure out various factors responsible for lower accuracy scores compared to best models. Proposed future work to improve the accuracy further we need to work with high resolution images and larger models(Efficient Net B5 or B6, Resnet 152 or 101). Use Ensemble learning for data imbalances.