A Deep Learning Based Classification of Abnormalities from Lung Ultrasound Images

Abstract

Lung Ultrasound (LUS) has become an increasingly valuable diagnostic tool in pulmonary imaging due to its non-invasive nature, portability, and ability to provide real-time imaging. Despite these advantages, several challenges persist in automated LUS analysis, including noise artefacts, low-contrast images, segmentation inaccuracies, and the lack of uncertainty quantification in AI-driven diagnostics. These limitations hinder the reliability of deep learning models in LUS-based disease detection and clinical decision-making. This research aims to address these challenges by developing a comprehensive framework that integrates advanced pre-processing, segmentation, uncertainty quantification, and classification techniques to enhance the accuracy and interpretability of LUS diagnostics. newlineTo improve image quality, this study introduces an Optimized Gaussian Histogram Equalization (OGHE) technique that enhances contrast while reducing noise artefacts. This pre-processing step ensures that input images maintain structural integrity, thereby improving segmentation accuracy and feature extraction reliability. The study further presents the Neoteric Edge-Based Segmentation Approach, which refines lung region segmentation by integrating edge detection, texture analysis, and gradient-based thresholding. This approach ensures the precise delineation of lung structures while minimizing the inclusion of irrelevant background regions, thereby enhancing diagnostic precision newline