Ensemble Deep Learning for the Detection of COVID-19 in Unbalanced Chest X-ray Dataset

Abstract

The ongoing COVID-19 pandemic has caused devastating effects on humanity worldwide. With practical advantages and wide accessibility, chest X-rays (CXRs) play vital roles in the diagnosis of COVID-19 and the evaluation of the extent of lung damages incurred by the virus. This study aimed to leverage deep-learning-based methods toward the automated classification of COVID-19 from normal and viral pneumonia on CXRs, and the identification of indicative regions of COVID-19 biomarkers. Initially, we preprocessed and segmented the lung regions usingDeepLabV3+ method, and subsequently cropped the lung regions. The cropped lung regions were used as inputs to several deep convolutional neural networks (CNNs) for the prediction of COVID-19. The dataset was highly unbalanced; the vast majority were normal images, with a small number of COVID-19 and pneumonia images. To remedy the unbalanced distribution and to avoid biased classification results, we applied five different approaches: (i) balancing the class using weighted loss; (ii) image augmentation to add more images to minority cases; (iii) the undersampling of majority classes; (iv) the oversampling of minority classes; and (v) a hybrid resampling approach of oversampling and undersampling. The best-performing methods from each approach were combined as the ensemble classifier using two voting strategies. Finally, we used the saliency map of CNNs to identify the indicative regions of COVID-19 biomarkers which are deemed useful for interpretability. The algorithms were evaluated using the largest publicly available COVID-19 dataset. An ensemble of the top five CNNs with image augmentation achieved the highest accuracy of 99.23% and area under curve (AUC) of 99.97%, surpassing the results of previous studies.

Mortality Prediction of COVID-19 Patients Using Radiomic and Neural Network Features Extracted from a Wide Chest X-ray Sample Size: A Robust Approach for Different Medical Imbalanced Scenarios

“…Deep learning (DL) algorithms have been extensively applied for COVID-19 detection/segmentation of infected pneumonia regions from HRCTs and CXRs [12][13][14][15][16]. Shiri [12] built a residual network to develop a fast, consistent, robust and human error immune framework for lung and pneumonia lesion detection and quantification.…”

“…In addition, it is well known that class imbalance is one of the main causes of the decrease of generalization in DL and ML algorithms [15,16,[30][31][32][33][34]. Bridge [16] proposed a novel activation function to improve COVID diagnosis performance when one class significantly outweighs the other.…”

Rapid diagnosis of Covid-19 infections by a progressively growing GAN and CNN optimisation

“…Since high resolution synthetic images are used in this study, with detailed features, the images have inherently less noise compared to low-resolution images and thus perform better even without segmentation. Furthermore, we achieve a better test accuracy than a standalone CNN network due to the limited Covid X-ray image dataset available for training [42] . Following this development of neural network architectures for generative learning, high resolution X-ray images are obtained, which are of equal realistic quality as the real images.…”

“…In [20] , the authors have utilized multiple image augmentation techniques such as random flipping, random jitter and random cropping to augment the input pipeline. Semi-automated and automated classification algorithms have been proposed in [39] , [42] to overcome this issue to classify unlabeled data. In addition to classical augmentation methods, Deep learning based augmentation techniques have been gaining a lot of momentum in recent years, especially Generative modelling based methods such as GANs [27] .…”

Generative adversarial network based data augmentation for CNN based detection of Covid-19

“…In 30 , the authors have utilized multiple image augmentation techniques such as random flipping, random jitter and random cropping to augment the input pipeline. Semi-automated and automated classification algorithms have been proposed in 31,32 to overcome this issue to classify unlabeled data. In addition to classical augmentation methods, Deep learning based augmentation techniques have been gaining a lot of momentum in recent years, especially Generative modelling based methods such as GANs 33 .…”