Robust Computer Vision for POCUS: Achieving Reliability in Ultrasound Imaging

This thesis addresses the critical need for interpretable AI in medicine. Within the TiAssisto telemedicine project, it proposes an efficient ensemble method for Lung Ultrasound (LUS) classification, achieving 100% accuracy on edge devices. To ensure clinical trust, the research introduces Intrinsically Guided Training (Batch-CAM). By embedding interpretability directly into the loss function, the model is actively constrained to learn clinically relevant features rather than spurious correlations, paving the way for transparent and trustworthy healthcare AI.