Computer-aided detection system for pulmonary embolism with integrated cardiac assessment based on embolic burden

Pulmonary embolism (PE) is a cardiovascular disease re- sulting from occlusion(s) in the pulmonary arteries. Its definitive diagnosis relies mainly on imaging, being comput- erized tomography pulmonary angiogram the gold standard. Recently, there has been increasing interest in automatiz- ing PE detection with the use of computer-aided detection systems, aiming to reduce workloads and enhance identifi- cation. Manual semiquantitative scores of embolic burden have also been proposed to assess PE severity and reinforce management. Yet, few attempts have been made to couple both. Here, we propose a deep learning-based system for PE detection, which exploits the visual explanations from the detector network to represent and quantize embolic burden. The resulting measurements of embolic burden are used to assess cardiac function, using a univariate logistic regres- sion model. Particularly, we propose to predict right-to-left ventricle diameter (RV/LV) ratio ≥1, a prognostic cardiac feature strongly associated with both embolic burden and ul- timate clinical outcome. The detector network is based on a Squeeze-and-Excitation-ResNet50 and trained on a subset of the RSNA-STR Pulmonary Embolism CT dataset. For the PE detection task, we achieve an accuracy of 0.72, sensitiv- ity of 0.73, and specificity of 0.82 on the test set, which is slightly below the performance of radiologists. As the cardiac assessment directly depends on the detector’s performance, we are currently unable to successfully predict RV/LV ratio ≥ 1. Nevertheless, we believe our system is theoretically feasible and could assist in both PE detection and severity assessment in the future.