Accelerating Covid-19 Differential Diagnosis with Explainable Ultrasound Image Analysis: An Ai Tool

ABSTRACT

Introduction Care during the COVID-19 pandemic hinges upon the existence of fast, safe, and highly sensitive diagnostic tools. Ultrasound has practical advantages over other radio-logical modalities and can serve as a globally-available first-line examination technique. However, the specific LUS patterns such as B-lines or subpleural irregularities can be hard to discern, calling into play AI-based image analysis as a support tool for physicians. Methods A LUS dataset of patients with COVID-19, bacterial pneumonias, non-COVID-19 viral pneumonia and healthy volunteers was constructed to assess the value of deep learning methods for the differential diagnosis of COVID-19. We hypothesized that a frame-based convolutional neural network would correctly classify COVID-19 LUS with a high sensitivity and specificity. Results 202 LUS videos were analysed. The frame-based convolutional neural network correctly classified COVID-19 with a sensitivity of 0.90 ± 0.08 and a specificity of 0.96 ± 0.04 (frame-based sensitivity 0.88 ± 0.07, specificity 0.94±0.05). We further employed class activation maps for the spatio-temporal localization of pulmonary biomarkers, which we subsequently validated for human-in-the-loop scenarios in a blindfolded study with medical experts. Aiming for scalability and robustness, we also performed ablation studies comparing mobile friendly, frame- A nd video-based architectures and show reliability of the best model by aleatoric and epistemic uncertainty estimates. We validated our model on an independent test dataset of 39 videos with COVID-19 severity scores and report promising performance (sensitivity 0.806, specificity 0.962). Figure 1 shows the flowchart. Conclusion Our work shows the potential of interpretable AI to serve as a decision support system for diagnosis and thereby provide an accessible and efficient screening method. Further clinical validation of the proposed method is underway. Data and code are publicly available at https//github. com/jannisborn/covid19-pocus-ultrasound.