Optimized Approach for Precise Segmentation of COVID-19 Infected Regions in Chest X-ray Images

Abstract

The swift and devastating impact of infectious diseases like COVID-19 has caused significant health and economic losses globally. Non-invasive methods, such as chest X-rays, are critical for COVID-19 detection, given the labor-intensive and time-consuming nature of PCR-based diagnosis. Advanced image processing techniques enhance chest X-ray analysis by improving image quality, extracting critical features, isolating lung regions, facilitating automated detection of infection patterns, and supporting radiologists in accurate diagnosis and disease monitoring. This paper introduces a hybrid methodology that leverages the UNet3+ model with a ResNet50 backbone for precise lung segmentation, ensuring enhanced focus and accuracy in subsequent analyses. The lung segmentation guides an iterative global thresholding (IGT)-based approach to detect COVID-19 infection patterns. Additionally, the methodology incorporates the Watershed algorithm as a post-processing step to refine the region of interest (ROI) boundaries, ensuring improved delineation of infection areas and reducing artifacts or noise. This integrated approach enhances the reliability and precision of automated COVID-19 diagnosis using chest X-rays. Experiments conducted on datasets from Lakeview Hospital, Goaves, Belagavi, and Kaggle validate the proposed method. Results show high performance metrics, including Accuracy (0.95), Precision (0.95), Recall (0.92), F1 Score (0.95), Dice Coefficient (0.97), Specificity (0.98), and Jaccard Index (0.96), with a mean squared error (MSE) of 0.1. The proposed method outperforms segmentation techniques like Adaptive Mean Thresholding, Otsu, Sauvola, and Niblack. Comparisons with manual results from medical experts and radiologists further confirm its effectiveness.