Dual Energy CT-enabled Asymptomatic Pulmonary Embolism Detection on Non-contrast CT

项目摘要 无症状性肺栓塞（PE）通常是通过对比计算机断层扫描偶然发现的 (CT)扫描不针对PE。它在患者中的平均患病率为2.6%， PE的死亡率和复发率。目前，放射科医生不阅读PE的非造影CT，因为 NCCT上溶栓高信号较弱。因此，约有2.6%的非典型肺炎患者可 无症状PE但根本未被诊断，这可能是一个大的人群。 我们提出了一种基于深度学习的单能量NCCT自动PE检测算法，以改善 从NCCT中发现无症状PE的成本效益。该算法将用于识别患有以下疾病的患者： PE的可能性较高，需要人工阅读或对比CT扫描。一个主要挑战是训练数据 由于无症状PE的患病率相对较低和阅读NCCT的难度，因此累积。到 为了克服这一挑战，我们建议使用双能CT（DECT），这是一种常规用于PE的方法 诊断，以生成虚拟非造影（VNC）图像作为训练图像。我们建议使用深度学习 用于VNC生成的算法，以填补VNC图像和真实的单能量NCCT之间的图像质量差距， 这确保了我们在VNC图像上训练的PE检测算法可以容易地应用于真实的NCCT。 该项目的预期成果是（1）深度学习算法，以生成逼真的VNC图像 对比DECT;（2）深度学习算法，以高灵敏度从NCCT中筛选PE。

Project Summary Asymptomatic pulmonary embolism (PE) are often incidentally discovered from contrast computed tomography (CT) scans that do not target PE. It has a mean prevalence of 2.6% among patients and associated with increased mortality rate and recurrence of PE. Currently non-contrast CT are not read by radiologists for PE, because the hyperintensity signal of thrombolysis on NCCT is weak. Hence, around 2.6% of the patients with NCCT can have asymptomatic PE but are not diagnosed at all, which is potentially a large population. We propose a deep learning-based automatic PE detection algorithm for single-energy NCCT to improve the cost-effectiveness to discover asymptomatic PE from NCCT. The algorithm will be used to identify patients with higher probability of PE and call for human reading or contrast CT scans. A major challenge is training data accumulation due to the relatively low prevalence of asymptomatic PE and hardness of reading NCCT. To overcome this challenge, we propose to utilize dual energy CT (DECT), which is becoming routinely used for PE diagnosis, to generate virtual non-contrast (VNC) images as training images. We propose to use deep learning algorithm for the VNC generation to fill the image quality gap between VNC images and real single-energy NCCT, which ensures that our PE detection algorithm trained on VNC images can be readily applied to real NCCT. The expected outcome of the project is (1) a deep learning algorithm to generate realistic VNC images from contrast DECT; (2) a deep learning algorithm to screen PE from NCCT with high sensitivity.