Efficacy of Transfer Learning-based ResNet models in Chest X-ray image classification for detecting COVID-19...

Because of COVID-19's effect on pulmonary tissues, Chest X-ray(CXR) and Computed Tomography (CT) images have become the preferred imaging modality for detecting COVID-19 infections at the early diagnosis stages, particularly when the symptoms are not specific. A significant fraction of individuals with COVID-19 have negative polymerase chain reaction (PCR) test results; therefore, imaging studies coupled with epidemiological, clinical, and laboratory data assist in the decision making. With the newer variants of COVID-19 emerging, the burden on diagnostic laboratories has increased manifold. Therefore, it is important to employ beyond laboratory measures to solve complex CXR image classification problems. One such tool is Convolutional Neural Network (CNN), one of the most dominant Deep Learning (DL) architectures. DL entails training a CNN for a task such as classification using extensive datasets. However, the labelled data for COVID-19 is scarce, proving to be a prime impediment to applying DL-assisted analysis. The available datasets are either scarce or too diversified to learn effective feature representations; therefore Transfer Learning (TL) approach is utilized. TL-based ResNet architecture has a powerful representational ability, making it popular in Computer Vision. The aim of this study is two-fold- firstly, to assess the performance of ResNet models for classifying Pneumonia cases from CXR images and secondly, to build a customized ResNet model and evaluate its contribution to the performance improvement. The global accuracies achieved by the five models i.e., ResNet18_v1, ResNet34_v1, ResNet50_v1, ResNet101_v1, ResNet152_v1 are 91.35%, 90.87%, 92.63%, 92.95%, and 92.95% respectively. ResNet50_v1 displayed the highest sensitivity of 97.18%, ResNet101_v1 showed the specificity of 94.02%, and ResNet18_v1 had the highest precision of 93.53%. The findings are encouraging, demonstrating the effectiveness of ResNet in the automatic detection of Pneumonia for COVID-19 diagnosis. The customized ResNet model presented in this study achieved 95% global accuracy, 95.65% precision, 92.74% specificity, and 95.9% sensitivity, thereby allowing a reliable analysis of CXR images to facilitate the clinical decision-making process. All simulations were carried in PyTorch utilizing Quadro 4000 GPU with Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60 ​GHz processor and 63.9 ​GB useable RAM.