Reliable and trustworty learning prototype: insight from POCUS

Deep learning models often lack the interpretability and trustworthiness required for clinical use. This paper proposes a prototype-regularised training method to analyse 1,208 lung ultrasound images, focusing on B-line artefacts. A ConvNeXt- Tiny architecture is used, adding a novel reconstruction loss to the standard classification loss. The model is guided to extract meaningful prototypes and uses them to classify the ultrasound images. To prevent these constraints from hindering generalisation, it is used in pairs with the proposed reconstruction loss, a set of plausible data augmentation of the ideal researched prototypes, and a geometry-aware network, a spatial transformer network, to measure which solutions help the network towards outputting the most reliable outcomes. The resulting models are precise, lightweight and interpretable, indicating that the proposed solution can be embedded in an ultrasound device to assist healthcare specialists in point-of-care applications.