Lw-IRBCaSa-Net: A Lightweight Inverted Residual Network with CaSa Attention Module for Disease Prediction from CXR Images

The coronavirus disease took a heavy toll on humans and threatened their existence. Millions died, leaving several recovered patients with severe long term side-effects. It is crucial to demarcate COVID-19 from Pneumonia as most of the symptoms in COVID-19 patients and Pneumonia patients are identical. The agenda is to classify patients’ chest X-ray (CXR) images into normal (or uninfected), COVID-19 Positive and Viral Pneumonia categories. This article proposes Lw-IRBCaSa-Net, an extremely Light Weight Depthwise Separable Convolutional Neural Network (DS-CNN) having a CaSa attention module integrated into an intermediate Inverted Residual Block (IRB) for COVID-19 and Pneumonia detection in CXR. Lw-IRBCaSa-Net comprises approximately 4 M parameters, which accounts for its compactness and lightweight attribute. Moreover, Lw-IRBCaSa-Net can achieve state-of-the-art F1-score of 95.08%. These results prove Lw-IRBCaSa-Net's superiority over other related works since it is lightweight and has no pre-trained bases. Additionally, we have proposed a novel attention mechanism known as the CaSa attention module, which plays a significant role in amplifying Lw-IRBCaSa-Net's performance.