In Journal of medical Internet research ; h5-index 88.0
BACKGROUND : COVID-19 has spread at a very fast rate and it is important to build a system that can detect it in order to help an overwhelmed health care system. The strength of deep learning techniques is used in many research studies of chest-related diseases. Although some of these researchers used state-of-the-art techniques and were able to provide promising results, these methods do not provide much benefit if they can only identify one type of disease without identifying the rest.
OBJECTIVE : The main purpose of this paper is a fast and more accurate diagnosis of COVID-19. This paper proposes a technique that classifies COVID-19 using X-rays from normal X-rays and X-rays of 14 other chest diseases.
METHODS : A novel multi-level pipeline was introduced based on deep learning models for the detection of COVID-19 along with other chest diseases from X-ray images. This pipeline reduces the burden of classifying a large number of classes on a single network. Deep learning models used were pre-trained models on ImageNet dataset and transfer learning was used for fast training. Lungs and heart are segmented from the whole X-ray image and passed onto the first classifier that checks if the X-ray is normal, COVID-19 affected or belongs to another chest X-ray disease. If the case is neither COVID-19 nor normal, then the second classifier comes into action and classifies the image as one of the other 14 diseases.
RESULTS : With our new pipeline, we show how our model makes use of the CNN state-of-the-art deep neural networks to achieve COVID-19 classification accuracy with 14 other chest diseases and normal X-ray images that are competitive with present state-of-the-art models. Due to lack of data in some classes like COVID-19, 10-fold cross-validation was applied. Our classification technique achieved average accuracy after 10-fold cross validation through the ResNet50 model with training accuracy of 96.04% and test accuracy of 92.52% for the first level of classification (3 classes). For the second level of classification (14 classes), we achieved maximum training accuracy of 88.52% and test accuracy of 66.634% using ResNet50. We also showed that when classifying all the 16 classes at once, overall accuracy decreased for detection of COVID-19 which in the case of ResNet50 was 88.92% for training data and 71.905% for test data.
CONCLUSIONS : This paper proposed a pipeline that detected COVID-19 with a higher accuracy along with 14 other chest diseases using X-rays. We showed that our pipeline provides better accuracy by dividing the classification task into multiple steps rather than classifying them collectively.
Albahli Saleh, Yar Ghulam Nabi Ahmad Hassan