Deep Learning Classification of Venous Thromboembolism Based on Ultrasound Imaging

Venous thromboembolism (VTE) occurs when a blood clot forms in a vein. According to the US National Institutes of Health, VTE affects 0.13% of men and around 0.11% of women in the United States every year, i.e., about 400 000 people per year. VTE includes deep vein thrombosis (DVT) and pulmonary embolism (PE). DVT is linked to the obstruction of a deep vein by a blood clot, usually in the lower leg, thigh, or pelvis. Whereas pulmonary embolism (PE) results from the migration of the blood clot toward a pulmonary artery. The objective of our project is to evaluate the possibility of predicting a PE based on ultrasound (US) images. It should be emphasized that there is no medical expertise for the detection of PE from these images. We proposed two methods: the first is based on the extraction of texture descriptors and the second relies on deep learning models. We developed a learning scheme for deep neural networks based on a joint training on a classification and segmentation task, and then a specialization of the network on the classification task. Alternatively, we built a model combining images and clinical data. Beyond the techniques used, significant work has been carried out to sort the database studied and select images. We obtained conclusive accuracy on the detection of PE.