Histogram Matching Based Data-Augmentation and Its Impact on CNN Model for Covid-19 and Pneumonia Detection from Radiology Images

In this study, we present a Histogram-Matching based novel data augmentation method for Covid-19 and Pneumonia detection, from an imbalanced Radiology (Chest X-Ray) image dataset. Moreover, we have publicly shared a large augmented CXR dataset on the Kaggle site ( https://www.kaggle.com/datasets/vibhuti25/histogram-matching-covid-cxr ), which is more balanced than the original dataset. For data augmentation, we have incorporated the Histogram Matching technique, CLAHE 0.5, and CLAHE 0.75, in order to increase the number of images in the dataset. The number of images in the augmented dataset per class is chosen by a statistical formula which is computed from the dataset itself. We have implemented two existing Convolutional Neural Network (CNN) models for Covid-19 detection i.e., Covid Net and Covid-Lite models, on the proposed augmented dataset. We have achieved 3–5% and 7–9% improvement of testing accuracy, precision, recall and F1 score for Covid Net and Covid-Lite model respectively, after employing the proposed augmented dataset. Furthermore, we have implemented a popular CNN model i.e., VGG-16 and by the proposed data-augmentation we have achieved the best results of 95% accuracy, 96% F1 score, 96% precision and 96% recall.