Developed a Convolutional Neural Network based on VGG16 architecture to diagnose COVID-19 and classify chest X-rays of patients suffering from COVID-19, Ground Glass Opacity and Viral Pneumonia. This repository contains the link to the dataset, python code for visualizing the obtained data and developing the model using Keras API.
Abstract—Today, Deep learning and Computer Vision are being leveraged by almost every industry across the globe. They have now become ubiquitous tools for research in the fields of Robotics, Autonomous cars, Computer vision, medical and health sciences to name a few. In this paper, we equip frontiers of Deep learning and Computer Vision to diagnose COVID-19. The primary clinical method currently in use for the diagnosis of COVID-19 is the Reverse Transcription Polymerase Chain Reaction (RT-PCR), which is expensive and requires trained medical personnel. Radiography is an easily accessible tool that can be a reasonable alternative to RT-PCR in diagnosing COVID-19. A Convolutional Neural Network based on VGG16 architecture is trained and analyzed on around 21000 lung X-ray images using transfer learning. Out of the 21170 images obtained from Kaggle repository, 16500 images have been used for training, 3130 have been used for validation and 1540 for testing the validated model. The goal is to accurately screen the patients suffering from Covid-19 against those who also suffer from Ground Glass Opacity and Viral Pneumonia which have a similar effect on human lungs as that of Covid-19. In the result analysis, the model gives a train accuracy of 99.16% and a validation accuracy of 98.56%. The proposed model helps radiologists diagnose COVID-19 within 0.5 seconds in a system equipped with GPU (Graphic Processing Unit) by classifying thousands or even millions of images in a single click. When trained with a larger dataset, the model may lead to facilitating early treatment of such lethal disease resulting in improved clinical outcomes. This work just proposes a possible method of screening COVID-19 infected patients and does not claim any medical accuracy.
DATASET The dataset for training, development and testing the model has been obtained from Kaggle repository [1] which comprises lung X-ray images (All the images are in Portable Network Graphics file format with a resolution of 299*299 pixels.) of patients infected with COVID-19, Ground Glass Opacity or Lung Opacity, Viral Pneumonia along with those of healthy people. The database was created by a team of researchers from Qatar University and University of Dhaka with an objective that researchers can use it to produce impactful work on COVID-19. In the second update, it consists of 21170 lung X-ray images out of which 3616 images are of COVID-19, 6012 images are of Ground Glass Opacity, 1345 images are of Viral Pneumonia and 10200 lung X-ray images of healthy people. This dataset is further divided into training, validation and testing sets each consisting of 16500, 3130 and 1540 images respectively so that the training set comprises 80% of the data obtained. It should be noted that no data cleansing is required in this case.
The model has been trained and validated on Google COLAB platform which is a very useful tool for executing deep learning algorithms as it gives free access to dedicated GPUs and TPUs (Tensor Processing Unit). The base hardware for the training face of the model is the TPU offered by Google over the cloud. A TPU is a hardware accelerator specialized for deep learning tasks. It shortens the training time by performing matrix multiplication in the hardware. At the core of a TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) as stated in [2]. TPU is on average about 15X-30X faster than its contemporary GPU (Nvidia K80) and CPU (Intel Haswell) with TOPS/Watt about 30X-80X higher. According to the authors, the matrix unit uses systolic execution (perfect fit for CNNs) to save energy and time by reducing reads and writes of the unified buffer. Subsequently, the validation and testing phases were carried out on Tesla K80 GPU over the cloud.
[1] Link to the dataset: https://www.kaggle.com/tawsifurrahman/covid19-radiography-database [2] Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, “In-Datacenter Performance Analysis of a Tensor Processing Unit”, N. P. Jouppi et al., ISCA 2017.