Crowdsourcing an Open COVID-19 Imaging Repository for AI Research

Purpose

The objectives of this project are to (1) assemble a crowdsourced, de-identified radiographic repository; and (2) train and validate existing COVID-NET deep learning diagnostic models.

Condition

  • COVID-19

Eligibility

Eligible Ages
Over 18 Years
Eligible Genders
All
Accepts Healthy Volunteers
Yes

Inclusion Criteria

  • This study will include anyone in the country who has been tested for COVID-19 with a chest radiograph image.

Exclusion Criteria

  • Patients who do not have a chest radiograph image used for COVID-19 testing

Study Design

Phase
Study Type
Observational [Patient Registry]
Observational Model
Other
Time Perspective
Other

Recruiting Locations

University of Central Florida
Orlando, Florida 32827
Contact:
Amoy Fraser, PhD
407-266-8742
amoy.fraser@ucf.edu

More Details

NCT ID
NCT05384912
Status
Recruiting
Sponsor
University of Central Florida

Study Contact

Amoy Fraser, PhD, CCRP, PMP
4072668742
amoy.fraser@ucf.edu

Detailed Description

The COVID-19 pandemic is laying bare the need for accessible curated datasets that researchers can use to build clinical-grade artificial intelligence (AI) models. Researchers in China recently used deep learning models of clinical-grade AI trained on radiographic imaging at an exponential scale to detect COVID-19 cases and optimize allocation of limited resources. (Jin S, Wang B, Xu H, et al. Ai-assisted ct imaging analysis for covid-19 screening: Building and deploying a medical ai system in four weeks. medRxiv. 2020:2020.2003.2019.20039354. doi: 10.1101/2020.03.19.20039354). This research platform is currently not possible in the United States because there are no large accessible radiographic image sets of COVID-19 patients to leverage. Therefore, the purpose of this project is to launch an interactive and HIPAA-compliant web portal-CovidImaging.com-where patients can securely share their radiographic imaging data. This portal will serve as an imaging repository for the purpose of training, testing, and validating an AI model aimed at earlier and more accurate disease detection in this global fight against COVID-19. On January 30, 2020 the World Health Organization designated the COVID-19 outbreak that originated in Wuhan, China as a global health emergency. Since then, the virus has rapidly spread across the world as a pandemic, unfavorably affecting health care systems at the expense of primary healthcare requirements.1Symptomatic cases of COVID-19 present with clinical symptoms similar to viral pneumonia such as fever, shortness of breath, chills, fatigue, cough, and dyspnea that can progress to acute respiratory distress syndrome, requiring critical care and ventilation. 2 Bronchoalveolar lavage analysis and electron microscopy identified the causative agent to be a novel, positive-sense RNA virus in the Coronaviridae family, with spiked peplomers attached to its envelope.3 This family of viruses has also been associated with severe acute respiratory syndrome (SARS) and Middle East respiratory syndrome (MERS), which cause similar pneumonia-related mortality. Preliminary reviews have been conducted to investigate the overlap of reported imaging features in SARS, MERS, and COVID-19 as it relates to onset of symptoms, progression of disease, and follow up. Early evidence suggests significant overlap in imaging features such as subpleural and peripheral areas of ground-glass opacity and consolidation, with initial chest imaging indicating abnormality in at least 85% of COVID-19 patients. 4 In the absence of vaccines and specific therapeutic drugs for the prevention and treatment of COVID-19, detection of the disease plays a vital role in containment strategies that isolate infected people from the healthy population. Even though RT-PCR sensitivity for COVID-19 can be as low as 60-70%, it is currently the large-scale method of testing with its high specificity.7 The low sensitivity of RT-PCR, along with limitations of sample collection, time delay, transportation, and lab equipment, means that not enough COVID-19 positive people are being identified in time to prevent progressive infection of this highly contagious virus. Given the respiratory involvement in COVID-19 infections, chest radiography has played an important role in screening, diagnosing, and developing treatment plans for patients with COVID-19-related pneumonia. Therefore, combining imaging with clinical and laboratory findings could facilitate the early diagnosis of COVID-19.5 Early detection would speed up treatment and allow for early patient isolation. This is essential for the implementation of public health surveillance, containment, and response for a highly communicable disease in which transmission can occur prior to onset of symptoms. Improving the precision of radiographic interpretation with AI models may improve detection rate and patient prognosis and thus help to reduce COVID-19 spread. As recently reported, chest CT demonstrates common radiographic features in almost all COVID-19 patients, including ground-glass 4 opacities, multifocal patchy consolidation, and/or interstitial changes with a peripheral distribution. 8,9 Studies have also been conducted to compare the efficacy and diagnostic value of chest CT to RT-PCR tests in COVID-19 cases. A case report of 1014 patients in China concluded that chest CT has a high sensitivity for diagnosis of COVID-19, with 60% to 93% of cases showing initial positive CT diagnosis prior to the initial positive RT-PCR results. 10 Another study with 51 patients having chest CT and RT-PCR assay within 3 days showed that the sensitivity of CT for COVID-19 infection was 98%, compared to 71% RT-PCR sensitivity.11 These studies further indicate the diagnostic value of chest radiographs alongside clinical and laboratory findings. This project will develop a large radiographic image repository which will be used to train and validate an AI deep learning model. This project necessarily involves not only designing and refining a deep learning model, but also curating a repository of donated chest radiographs that will be used to train the novel model. Using a secure and HIPAA-complaint online platform, as has been designed for this project, will allow this project to employ a big data approach to improve the accuracy of the model. Since patients from healthcare facilities around the country will have equal opportunity to participate in the project, this portal will also provide an opportunity to expand the demographic pool in a way that previous studies could not.