Detection of Pulmonary Embolisms in NCCT Data Using nnDetection

Linda Vorberg, Florian Thamm, Hendrik Ditt, Marius Horger, Florian Hagen, Andreas K. Maier

Abstract: Pulmonary embolism (PE) is a life-threatening disease for which a prompt diagnosis is important and challenging as its symptoms are mostly nonspecific. In the clinical workflow of diagnosing PE, computed tomographic pulmonary angiography (CTPA) has become the gold standard imaging tool. As the performance of a CT scan with contrast agent sometimes can be contraindicated and is associated with high costs, identifying the embolism with a non-contrast CT (NCCT) scan is desirable. We automated the detection of PE in NCCT scans with the use of deep learning, in order to guide the physician and speed up the clinical workflow. We used nnDetection, designed for medical object detection, to accomplish this task. nnDetection gets informed by additional channels which are used besides the NCCT scan. These are segmentation masks of the lung lobes and vessels which are also used for post-processing. In a study with 99 patients that all presented with PE, nnDetection was shown to detect a PE in 71% of the cases when considering the first 10 boxes with the highest probability containing a PE.