Novel Machine Learning Methods for Lung Ultrasound Image Analysis

Abstract

Lung ultrasound (LUS) has emerged as a valuable imaging modality for assessing pulmonary conditions and guiding fluid management in dialysis patients. However, the manual interpretation of LUS images is time-consuming, operator-dependent, and lacks standardization. This thesis explores advanced machine learning techniques to automate LUS-based analysis, enhancing the accuracy, efficiency, and clinical applicability of B-line detection and fluid overload estimation.

Three key contributions are presented. First, a self-supervised learning approach for B-line detection is proposed, leveraging contrastive learning and fine-tuning with Faster-RCNN. This method reduces reliance on large annotated datasets and improves detection robustness compared to traditional model-based approaches. Second, deep unfolding networks are introduced for line artifact detection, bridging iterative optimization with neural networks. DUBLINE, based on alternating direction method of multipliers (ADMM) unfolding, enhances computational efficiency, while CPSNet, utilizing Cauchy proximal splitting (CPS), introduces an unsupervised loss function, Radon-Based Neighbor Reconstruction Loss, to eliminate dependence on ground truth labels. CPSNet achieves stable, real-time B-line detection with improved robustness against noise. Third, a multimodal learning framework is developed for fluid overload estimation in dialysis patients, fusing visual and tabular features. The analysis reveals the limitations imposed by dataset characteristics, including feature variability and data imbalance, which may constrain the benefits of integrating imaging and clinical data. Despite these challenges, this exploratory work lays the foundation for future research, emphasizing the need for richer physiological datasets, improved feature fusion techniques, and longitudinal modeling strategies to enhance predictive reliability.

By integrating self-supervised learning, deep unfolding, and multimodal fusion, this thesis contributes to the development of AI-assisted LUS analysis, paving the way for more accurate, scalable, and clinically deployable solutions for pulmonary assessment and dialysis management.

Date of Award	30 Sept 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Alin Achim (Supervisor), Pui Anantrasirichai (Supervisor) & Oktay Karakus (Supervisor)