GAN and CNN Model for Generation of Synthetic Medical Images and Classification

Authors:

Abstract
for any machine learning or deep learning model, data is the most essential ingredient. Data augmentation has become the need of the hour as the required amount of data is not available in many fields and real world scenarios. In such situations, synthetic data comes into the picture. Synthetic data generation techniques have made it possible to create a huge amount of data from really small data. It has been possible to make realistic artificial data that resembles real data. Finding a particular dataset that has required features is a difficult task. Medical data is rare and expensive. It is nearly impossible to get real data with clinical trials in the required amount. In this paper, we will be discussing, synthetic medical image generation with Generative Adversarial Networks (GANs). We used Deep Convolutional GAN for the generation of medical images. We were able to generate 400 synthetic images of retinal fundus from a tiny dataset of 20 images. We calculated the FID score to evaluate the generated synthetic images and we were able to keep it as low as 38.31. We applied the generated synthetic images and real images on CNN model for a comparative study on accuracies achieved with both the images. We were able to get the accuracy of 60% with synthetic images out of dataset consisting of only 20 images.