Imaging Analysis and Deep Learning based approaches for Automated Lung Ultrasound Pattern Classification

Since the outbreak of the COVID-19 pandemic, lung ultrasound (LUS) has been considered as a promising tool to assess patients' condition among clinicians worldwide. However, LUS assessment mainly relies on the interpretation of visual patterns. These patterns are visualized as hyper-echoic horizontal and vertical artifacts and hypo-echoic small to large consolidations. To assist clinicians in their analysis, various AI-based methods have been proposed. However, no extensive work has been done in spatial and temporal compression of LUS data to favor automated low-complexity LUS analysis, while retaining its clinical value. Assessment of these visual patterns using DL models in a real-world scenario, across different medical centers and different populations, is also an area worth investigating. Most importantly, a thorough analysis of state-of-the-art methods for the assessment of LUS patterns at the frame and video level, along with an extensive evaluation of these techniques in providing a clinical value of the analysis is yet missing in the literature. In this regard, this thesis addresses these challenges by proposing studies showing how down-sampling of LUS frames 32 times of their original size and compressing a LUS video to a single frame can allow achieving state-of-the-art performance at video and prognostic level. The thesis also discusses the analysis of the generalization capability of DL models when trained on data from different medical centers and tested on others. Going a step further, it also presents how well the DL models generalize across different geographic populations, age groups, context of pathology, acquisition protocol, and scoring systems. In this regard, DL models are trained on LUS data from adults and fine-tuned and evaluated on LUS data from children, setting a baseline for accurate LUS analysis in children. Lastly, a thorough benchmark analysis of AI-based techniques for LUS evaluation in the adult population affected by COVID-19-induced pneumonia, from frame to prognostic level is a worthy contribution to the domain knowledge.