RAPID: Collaborative Research: Independent Component Analysis Inspired Statistical Neural Networks for 3D CT Scan Based Edge Screening of COVID-19

点击翻译按钮获取中文摘要

COVID-19, the disease caused by the new coronavirus SARS-CoV-2, has shut down cities in the United State and around the world. Due to the global lack of test kits used to diagnose the disease, it is critical to screen suspected patients first and prioritize those most likely to have COVID-19 for further diagnostic test. As most patients with COVID-19 show visual signs of the pneumonia on images from chest Computerized Tomography (CT) scans, it is possible to screen patients based on these images. However, with the large number of suspected cases and the time required to analyze 3D images, radiologists are challenged to adequately screen all of the images. Most recently, several works have demonstrated the potential of deep neural networks in identifying typical signs or partial signs of COVID-19 pneumonia, drastically speeding up the screening process and reducing the burden on radiologists. Due to the large 3D volumetric data associated with chest CT scans (a few hundred MB per image), however, the deep neural networks for classification, which mostly work on 2D images only, do not work very well on 3D CT images. In this project, , the team explores novel solutions across software and hardware layers to enable a solution that allows plug-and-play for automatic COVID-19 screening with fast turn-around time. The project will enable the deployment of deep learning to efficiently and accurately screen suspected COVID-19 patients, and significantly reduce the burden on radiologists. It can effectively address the diagnosis bottleneck caused by the lack of rRT-PCR test kits. In addition, the proposed techniques can be applied to other areas beyond COVID-19 screening where neural networks need to handle large volumetric data. The project will be made open source to enable wide distribution in a timely manner.The proposed research will explore ICA-Net, a novel Independent Component Analysis (ICA) inspired statistical neural architecture that can efficiently and accurately extract features from 3D CT images of large sizes for COVID-19 screening. ICA-Net will be the first neural architecture that targets large volumetric 3D image classification. In addition, considering the practical use of this project where security/privacy of patient data and fast turn-around time are strongly desired, through hardware/software co-design, the project will identify the best solution to be deployed on the edge using commercially off-the-shelf hardware for plug-and-play in clinics. As such, it can be immediately integrated and used for COVID-19 screening.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.