RAPID: A Computational Deep-Learning Approach for Fast, Accurate CT Testing and Monitoring of COVID-19

冠状病毒病新冠肺炎现在是一场全球大流行，造成了前所未有的巨大健康和经济危机。尽管出现了新的检测方式，但仍然迫切需要快速、准确和可获得的工具来检测疑似新冠肺炎患者并监测疾病进展。ComputeCOVID19+项目通过提供一种基于计算的筛查工具满足了这一需求，该工具提供了比当前基于实验室的技术(即聚合酶链式反应)更高的筛查和监测准确性。ComputeCOVID19+系统还将使计算机断层扫描(CT)扫描的分析速度更快，以减轻放射科医生和医疗保健系统的负担。ComputeCOVID19+项目解决了冠状病毒筛查和监测方面的挑战：(1)从传统CT扫描仪重建超分辨率医学图像，(2)开发用于高保真图像重建和高精度新冠肺炎解释的新算法和软件，以及(3)用新冠肺炎临床数据验证我们的方法。该方法使用CT扫描和团队的超分辨率和基于去模糊的迭代重建(SAIDR)算法。结果表明，基于SADIR的神经网络具有更好的解释能力和鲁棒性。此外，它涉及的训练参数数量要少得多，因此更容易训练。最后，在网络训练中，萨迪尔不需要任何高分辨率的CT图像作为“地面真相”的参考。预期的结果是一种计算性的深度学习方法，可以检测和诊断新冠肺炎，具有高敏感性和高特异性。该方法还将能够更准确地监控新冠肺炎的疾病进展。该奖项反映了美国国家科学基金会的法定使命，并已通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The coronavirus disease COVID-19 is now a global pandemic, causing a huge health and economic crisis at an unprecedented scale. Despite new testing modalities, there remains an urgent need for fast, accurate, and accessible tools to test people of suspected COVID-19 and monitor disease progression. The ComputeCOVID19+ project addresses this need by providing a computationally-based screening tool that delivers much higher accuracy for screening and monitoring than current laboratory-based technique (i.e., PCR). The ComputeCOVID19+ system will also make analysis of Computerized Tomography (CT) scans faster to reduce the burden on radiologists and healthcare systems. The ComputeCOVID19+ project addresses the challenges of COVID screening and monitoring in (1) reconstructing super-resolution medical images from conventional CT scanners, (2) developing novel algorithms and software for high-fidelity image reconstruction and high-precision interpretation of COVID-19, and (3) validating our approach with clinical COVID-19 data. The method uses CT scans and the team’s super-resolution and deblur-based iterative reconstruction (SADIR) algorithm. As a result, the SADIR-based neural network has better explanation and robustness. In addition, it involves a much smaller number of training parameters, and hence, is easier to train. Finally, SADIR does not require any high-resolution CT images as the “ground truth” reference during network training. The expected outcome is a computational deep learning method that can detect and diagnose COVID-19 with high sensitivity and high specificity. The method will also enable monitoring of COVID-19 disease progression with better accuracy.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.