Fusion of images and clinical features for the prediction of Pulmonary embolism in Ultrasound imaging

Abstract

Venous Thromboembolism (VTE) is a life-threatening disease encompassing pulmonary embolism and deep venous thrombosis (DVT). Pulmonary embolism occurs in 50% of patients with a proximal deep venous thrombosis. We aimed to predict the occurrence of a pulmonary embolism in patients with a DVT from clinical data and Ultrasound images of proximal thrombosis. To address this task, we proposed to use a Deep learning model that uses both images and 5 clinical factors as input and we aimed to measure the contributions compared to using only images. Promising results were obtained with both models compared to the state-of-art. The contribution of the clinical factors remains unclear but a gain in accuracy was observed when using smaller models.