Advancing AI-driven lung ultrasound diagnostics for COVID-19: Procedural data synthesis, image enhancement, and pleural line detection

Abstract

This thesis presents novel advancements in lung ultrasound imaging through the integration of computational techniques and artificial intelligence. Addressing critical challenges such as low contrast, noise, data scarcity, and diagnostic variability, the research focuses on three core areas: contrast enhancement, synthetic data generation, and AI-driven diagnostic models for COVID-19 detection. A contrast enhancement technique utilising Rayleigh Gaussian Mixture Models and k-means clustering was developed to improve image clarity while preserving diagnostic features. This method demonstrated substantial improvements in pleural line detection accuracy, with Support Vector Machine (SVM)-based models achieving 84.3% accuracy and superior preci-sion metrics. To mitigate the scarcity of annotated lung ultrasound datasets, synthetic image gener-ation methods were implemented. Generative Adversarial Networks (GANs) were used to produce realistic ultrasound images, achieving a Structural Similarity Index (SSIM) of 0.46 and a Fr´echet Inception Distance (FID) of 257.95, showcasing their potential in addressing data limitations. The thesis also introduces P-Net, a multi-architecture neural network ensemble, designed to classify lung ultrasound images for COVID-19 diagnostics. P-Net integrates segmentation outputs and latent features, achieving a Dice coefficient of 0.87 and Intersection-over-Union (IoU) of 0.89, demonstrating robust diagnostic accuracy even in limited data scenarios. These contributions enhance the utility of lung ultrasound imaging, particularly for diag-nosing COVID-19, by combining advanced computational methods with clinical needs. This research underscores the transformative potential of artificial intelligence in medical imaging, paving the way for improved patient outcomes and innovations in healthcare technology.