LWSE: a lightweight stacked ensemble model for accurate detection of multiple chest infectious diseases including COVID-19

Sohaib Asif, Ming Zhao, Fengxiao Tang, Yusen Zhu

Abstract: Recently, the COVID-19 disease has significantly impacted global economies and healthcare systems. Swift and accurate detection of COVID-19 is crucial for effectively mitigating the spread of this pandemic. Chest X-ray images (CXR) and CT scans have emerged as valuable diagnostic tools for COVID-19 patients. However, existing deep learning (DL) methods for COVID-19 detection are often computationally expensive and require substantial memory resources. Therefore, there is a pressing need for a lightweight and computationally efficient solution to facilitate COVID-19 detection. In response to these challenges, we propose an innovative and efficient lightweight stacked ensemble model, known as LWSE. Our approach combines the MobileNet model with a lightweight convolutional neural network (CNN) to enhance the detection performance of various chest infectious diseases, including COVID-19. The integration of these models not only improves the learning capability but also significantly reduces the computational complexity. The stacked ensemble technique is employed to aggregate predictions from the MobileNet and lightweight CNN, leading to enhanced detection accuracy. Subsequently, these predictions are fused into a multilayer perceptron (MLP) for classification, yielding superior results compared to using a single model. To evaluate the effectiveness of our LWSE model, we have developed a novel COVID-19 dataset comprising 900 CXR images of Pakistani patients collected from local hospitals. Additionally, we conducted experiments on publicly available datasets. Our comprehensive evaluation benchmarks the LWSE model against four pre-trained models in terms of computational cost and detection performance. Notably, our LWSE model achieves highly promising results with an accuracy of 96.40% and 97.89% on the CXR dataset, and an outstanding accuracy of 98.83% on the CT dataset. The superior performance of our LWSE model, coupled with its low computational cost, enables faster image classification compared to pre-trained and state-of-the-art models. Consequently, our proposed LWSE model presents a more viable solution for rapid diagnosis of patients with chest infections.