Accelerating Detection of Lung Pathologies with Explainable Ultrasound Image Analysis

Care during the COVID-19 pandemic hinges upon existence of fast, safe, and highly sensitive diagnostic tools. Considering significant practical advantages lung ultrasound (LUS) over other imaging techniques, but difficulties for doctors in pattern recognition, we aim to leverage machine learning toward guiding diagnosis from LUS. We release largest publicly available LUS dataset consisting 202 videos four classes (COVID-19, bacterial pneumonia, non-COVID-19 viral pneumonia healthy controls). On this dataset, perform an in-depth study value deep methods differential pathologies. propose a frame-based model that correctly distinguishes data with sensitivity 0.90±0.08 specificity 0.96±0.04. To investigate utility proposed method, employ interpretability spatio-temporal localization pulmonary biomarkers, which are deemed useful human-in-the-loop scenarios blinded medical experts. Aiming robustness, uncertainty estimation demonstrate recognize low-confidence situations also improves performance. Lastly, validated our on independent test report promising performance (sensitivity 0.806, 0.962). The provided facilitates validation related methodology community framework might aid development accessible screening method diseases. Dataset all code at: https://github.com/BorgwardtLab/covid19_ultrasound.