DSENetk: An Efficient Deep Stacking Ensemble Approach for COVID-19 Induced Pneumonia Prediction Using Radiograph Images

Sandeep Chaurasia, Amit Kumar Gupta, Poorvi Pankaj Tiwari, Ananya Smirti, Mahesh Jangid

Abstract: Eruption of pandemic disease in 2019 has been spreading like a wildfire throughout the world, having approximately 237,222,557 cases according to the Worldometers. In many instances, the virus is not detectable using the Reverse Transcription Polymerase Chain Reaction (RT-PCR) test but can be identified through chest radiograph images. Thus, an automated system capable of early symptom prognosis and aiding medical professionals in the swift identification of infected patients is crucial. This study introduces a novel approach that significantly enhances the accuracy of COVID-19 detection by integrating advanced image enhancement techniques—Contrast Limited Adaptive Histogram Equalization (CLAHE) and Balance Contrast Enhancement Technique (BCET)—with an ensemble learning model comprising multiple Convolutional Neural Networks (CNNs), including Inception V3, ResNet50, Xception, and VGG16. The proposed system focuses on ternary classification, distinguishing between normal (healthy), COVID-19, and pneumonia cases, and is tested on a dataset of 7045 images, larger than those used in previous studies. Performance evaluation shows that the combined use of CLAHE and BCET led to a significant increase in accuracy, with the highest accuracy achieved by the stacked DSENetk (Xception + ResNet50) model at 95%. This model outperformed individual models, which had accuracies of 94% with CLAHE only and 92% with BCET only. The proposed DSENetk model also demonstrated superior performance compared to other existing models, making it a highly effective tool for COVID-19 detection through chest X-rays. These results underscore the potential of enhanced image preprocessing and ensemble learning for more accurate and reliable COVID-19 diagnosis.

External IDs:dblp:journals/sncs/ChaurasiaGTSJ25