Custom Weighted Balanced Loss function for Covid 19 Detection from an Imbalanced CXR Dataset

Abstract

In this paper, we have proposed a novel framework, that is ResNet-18 model along with Custom Weighted Balanced loss function, in order to automatically detect Covid-19 disease from a highly imbalanced Chest X-Ray (CXR) dataset. Covid 19 disease has become a global pandemic, for last two years. Early automatic detection of Covid-19, from CXR images has been the key to survive from this pandemic. In the recent advent, researchers have already proposed several Deep Learning (DL) models, which can detect Covid-19 disease (with higher accuracy) from CXR images. However, Covid-19 detection by DL models are fraught with the problem of class imbalance, since most of the available CXR datasets are found highly imbalanced. In this paper, we have worked in a new direction, that is, alleviating the class imbalance problem from CXR dataset by using novel loss function. First, we choose a challengeable CXR dataset in which there are four classes, they are Covid, Normal, Lung Opacity (LO) and Viral Pneumonia (VP). Later we have identified that real problem of this dataset is not only the class imbalance, but also, huge intra-class variance is observed in Covid class. Therefore, we have come up with a new idea, that is, modifying the bias weights in a Weighted Categorical Cross Entropy (WCCE), based on reducing both of the factors, i.e., class imbalance and intra-class variance from the dataset. For the experimentation, we have chosen a ResNet-18 model which is trained from scratch for a large Chexpert CXR dataset and thereafter it is pre-trained on the Covid CXR dataset. Experimental results suggest that ResNet-18 model along with proposed Custom Weighted Balanced loss function, have improved 2-4% accuracy, precision, recall, F1 score and AUC for four class CXR dataset. Furthermore, we have tested the same framework for three class Covid CXR dataset, after excluding LO class. We have achieved 96% accuracy, 97% precision, 96% recall, 97% F1 score and 97% AUC for three class classification task. This is significant (3-4%) improvement than the performance of ResNet-18 model with CCE.