A Lightweight COVID-19 predictive model with Synthetic CT images using Conditional GAN Knowledge Distillation

ABSTRACT

With the increase in the cases of COVID-19, the necessity of improving testing and treatment is increasing rapidly. Many techniques are currently being used by the medical fraternity for detection of COVID-19 in a patient such as RT-PCR, Chest CT Scan Images, Chest X-Ray scans, etc. Among these techniques, a Chest CT scan has proven to be highly accurate for screening of the novel coronavirus. But a trained professional like a radiologist is needed to analyze the CT scan and determine whether the patient is positive or not. Due to the sudden spike in the number of infections, there is a shortage of such professionals. A machine learning based system can be highly effective in assisting the doctors if it can accurately predict COVID-19 from a chest CT scan. However, the number of chest CT scan images available are very less in order to build an accurate machine learning based predictive model. We present a generative model for data augmentation of COVID-19 positive and negative Chest CT images. We use Conditional DCGAN for generating nearly 1502 COVID-19 positive and 1510 negative images thus extending a publicly available dataset. We also build predictive models using pre-trained models like VGG and ResNet to detect COVID-19, achieving an accuracy upto 87.7%. We also apply the technique of knowledge distillation to build a lightweight and computationally cheap predictive model that has an accuracy of 86.2% and is nearly 11 times smaller than the best model available on the dataset. © 2022 IEEE.