Abstract

Introduction: The medical field can utilize radiological images with deep learning techniques to diagnose disease more accurately, enabling the diagnosis and classification of a variety of illnesses. In the domain of learning and machine vision, identifying COVID-19 from Xray images is a developing area. Since the onset of COVID-19, significant work has been performed, yet some issues remain in this field.

Method: Firstly, there are limited X-ray scans readily available that are classified as COVID-19 positive, resulting in an unbalanced dataset. Secondly, there is no single set of data, classes, or evaluation protocols for all the work performed. This study proposes a three-class balanced dataset based on two validated publicly available datasets. Deep Convolutional neural networks have the potential to operate with both wide breadth and wide depth, which could raise computing complexity. Additionally, to deal with this issue, an attention-guided ensemble model (AGEM) is proposed to classify normal, pneumonia, and COVID-19 images. First, we propose an Attention Guided-Convolutional Neural Network (AG-CNN) architecture based on transfer learning. We used three pre-trained models i.e., InceptionV3, DenseNet121, and MobileNetV2, as the basis for the proposed AG-CNN, resulting in three attention-guided network architectures i.e., AGInceptionV3, AG-DenseNet121, and AG-MobileNetV2. Then, we used entropy computation and an uncertainty-based weighting ensemble to classify the images into three classes.

Result: The performance was evaluated and compared with existing works and 7 pre-trained models i.e., ResNet50, InceptionV3, VGG-16, VGG-19, Densenet-201, Xception, MobileNetV2, on our three-class dataset. An accuracy of 97.35%, recall of 97.35%, specificity of 98.67%, precision of 97.35%, and F1-score of 97.35% demonstrate the superiority of our proposed attention-guided ensemble model over pre-trained models and other existing studies.

Conclusion: It is noteworthy that for additional analysis, we utilized Grad-CAM or gradientweighted Class Activation Mapping.

Keywords: Attention mechanism, ensemble, CNN, transfer learning, entropy, uncertainty weighting, COVID-19, pneumonia