An explainable lightweight parallel depth-wise separable model for lung infection detection from chest X-rays

Respiratory diseases including severe acute respiratory syndrome (SARS), Middle East respiratory syndrome (MERS), HIV, influenza A (H1N1), and COVID-19 can cause lung infections. Among these, COVID-19 has been considered as an epidemic in the past few years. According to the World Health Organization (WHO), about 2.7 million deaths occurred due to lung infection caused by COVID-19. Therefore, most of the countries implemented full or partial lockdown measures to slow down the spread of the disease. However, early detection of viral infection is crucial to reduce the spreading as well as minimizing the pandemic situation. Common lung infection detection methods include X-ray image analysis, reverse transcription–polymerase chain reaction (RT-PCR), and computed tomography (CT) scans. RT-PCR and CT scans have limitations such as being error-prone, slow, and requiring specialized machinery. CT scans are expensive and expose patients to radiation. Similarly, manual assessment of X-ray images is challenging due to the limited availability of skilled medical professionals. To mitigate this problem, convolutional neural networks (CNNs) have shown potential in automating the detection of lung infections from X-ray images. However, the traditional CNN models have a large number of parameters and require high computational resources. To address these issues, this research proposes a novel parallel lightweight diagnosis model based on depth-wise separable CNN (LW-PDS-CovidNet). Additionally, the proposed model’s explainability is enhanced by utilizing Shapley additive explanations (SHAP) and gradient-weighted class activation mapping (GRAD CAM) to highlight the most important features. The proposed LW-PDS-CovidNet approach was tested on real chest X-ray images to predict them as normal, COVID-19, viral pneumonia, and lung opacity. Experimental outcomes demonstrate that the proposed method outperforms the baseline methods and achieved an accuracy of 98.06, 97.43, and 92% for two-class, three-class, and four-class classification respectively with a low number of parameters. Notably, the proposed model achieves these results with only 0.53 million parameters, whereas traditional CNN models require nearly 2.7 million parameters. This ability of the proposed technique makes it suitable for low-computing devices. Therefore, it is capable of significantly reducing cost and memory requirements for computing resources.