Abstract

Point-of-Care Ultrasound (POCUS) is a portable, cost-effective imaging modality with strong potential to expand access to diagnostic tools in remote and underserved settings. However, its interpretation still depends heavily on expert knowledge, which limits broader clinical adoption. This thesis aims to enhance the interpretability, reliability, and accessibility of POCUS by leveraging deep learning techniques.
The first part of this work presents a Bayesian deep learning framework for the classification of Non-alcoholic fatty liver disease using QUS features extracted from pre-clinical duck experiments. The model not only achieves accurate classification but also provides meaningful uncertainty estimates, helping assess prediction confidence. In the second part, we propose a method to decompose predictive uncertainty into epistemic and aleatoric components in the estimation of Homodyned-K distribution QUS parameters and investigate their relationship with prediction error. The final part introduces a multimodal dataset for the diagnosis of breast cancer-related lymphedema (BCRL) using POCUS. A deep learning pipeline is developed that integrates ultrasound images and clinical features to improve diagnostic performance.
Together, these contributions apply deep learning methods to enhance quantitative tissue characterization, uncertainty estimation, and diagnosis, making ultrasound more practical and accessible in everyday healthcare.