Automatic Detection and Multi-Class Classification of COVID-19, Pneumonia, and Tuberculosis Diseases in Chest X-ray Images Using Deep Learning Techniques

Abstract

Recent technology advancements have set the stage for deep learning-based approaches to be applied in nearly every aspect of life. The precision of deep learning techniques makes it feasible to be utilized in the medical field for the automatic detection and classification of various illnesses. The recent coronavirus (COVID-19) epidemic has put enormous strain on the global health system. COVID-19 can be diagnosed via PCR testing and medical imaging. Because COVID-19 is extremely infectious, and PCR tests has some limitations regarding the processing time and highly false positive probability, chest X-ray diagnosis is deemed safe in a variety of conditions as it has less radiation compared to CT-scans, along with being widely available and cost effective. A deep learning-based approach is suggested in this study to classify COVID-19 infection from other non-COVID-19 infections in the lungs. Multiclassification and detection of COVID-19, Normal, Pneumonia, and Tuberculosis was held using three distinct pre-trained models EfficientNetB0, Xception, and NasNetLarge which were selected based on their variance in size and accuracy on ImageNet. A dataset consists of 7135 X-ray images with four classes of Normal and different lung diseases including Covid-19 was utilized to train and evaluate the deep learning models, and two different training strategies were adopted. Not only three deep learning models were used in this study after tuning the optimum hyperparameters, but also a new classification head was generalized to boost performance and fit the classification problem along with data preprocessing techniques which were employed on the dataset such as pixel normalization and data augmentation to address the limitation of available data. Furthermore, during the model's training phase, a strategy with a class weight balancing technique was used, which significantly enhanced the performance of the deep learning models. Several performance parameters were used to assess the performance of the suggested strategies. NasNetLarge surpassed the other models especially in strategy II where class balancing technique was deployed in training phase with overall accuracy of 91.3%, recall, precision and F1 score of 0.94, 0.91, 0.91 respectively. EfficientNetB0 had also shown a competitive result with an accuracy of 88% since it has lesser execution time and considered more cost effective due its lightweight. This study is anticipated to be useful in enhancing medical practitioners' judgments and accuracy when identifying COVID-19 and other lung illnesses. This discovery will help future researchers reduce analysis duplication and pick the best network for their jobs.