DEEP NEURAL NETWORK CLASSIFICATION MODELS FOR COVID-19 DETECTION IN X-RAY IMAGES: TOWARDS FEW-SHOT AND META PSEUDO LABELS LEARNING

Abstract

Since the beginning of COVID-19 pandemic, efficient methods to detect the infection is in urgent need around the globe. There are existing testing systems created to handle the spread of infection. However, considering the fact that those systems can require substantial amount of time, and moreover, are not always commonly available, alternative testing methods have become necessary. In this study, our proposal is to use Convolutional Neural Networks (CNN) to detect COVID-19 infection appearance on chest X-ray images. During this study, we have utilized 9 different datasets that have been used to train and evaluate 5 CNN algorithms. The datasets contain X-ray images labeled as COVID-19, Pneumonia or Normal. CNN algorithms that have been used to classify the data include ResNet-50, VGG-16, AlexNet, Inception- V3, InceptionResNet-V2. Empirically, we have established that InceptionResNet-V2 model provides best evaluation accuracy averagely reaching accuracy of 95.1% which is a very promising result considering the medical nature of the domain used. To train models with a limited amount of data, we decided to use a new Few-Shot Learning method, which was able to achieve a result of 97.7%. We have also used a semisupervised learning method Meta Pseudo Labels, which allowed us to train models with a poor labeled datasets. The approach has also demonstrated promising results achieving 92.5% of accuracy on the data labeled only for 16%. Our performance results have made it possible to produce a heat map of an X-ray image that illustrates lung areas that are most influential for a model to distinguish COVID-19 images from images labeled as Normal. Additionally, we have provided 2-dimensional T-SNE representation that illustrates how the CNN models observe the data in lower dimension and separate it into clusters.