Transfer Learning based Detection of Diabetic Retinopathy from Small\n Dataset

Abstract

Annotated training data insufficiency remains to be one of the challenges of\napplying deep learning in medical data classification problems. Transfer\nlearning from an already trained deep convolutional network can be used to\nreduce the cost of training from scratch and to train with small training data\nfor deep learning. This raises the question of whether we can use transfer\nlearning to overcome the training data insufficiency problem in deep learning\nbased medical data classifications. Deep convolutional networks have been\nachieving high performance results on the ImageNet Large Scale Visual\nRecognition Competition (ILSVRC) image classification challenge. One example is\nthe Inception-V3 model that was the first runner up on the ILSVRC 2015\nchallenge. Inception modules that help to extract different sized features of\ninput images in one level of convolution are the unique features of the\nInception-V3. In this work, we have used a pretrained Inception-V3 model to\ntake advantage of its Inception modules for Diabetic Retinopathy detection. In\norder to tackle the labelled data insufficiency problem, we sub-sampled a\nsmaller version of the Kaggle Diabetic Retinopathy classification challenge\ndataset for model training, and tested the model's accuracy on a previously\nunseen data subset. Our technique could be used in other deep learning based\nmedical image classification problems facing the challenge of labeled training\ndata insufficiency.\n