AI + Lung US

Artificial intelligence in lung ultrasound in covid-19

Background

Infection with Covid-19 in adults can lead to severe hypoxemic respiratory failure that resembles acute respiratory distress syndrome (ARDS). Diagnosis and monitoring of this viral illness are difficult without serological testing; advanced imaging modalities like computed tomography (CT), while accurate, must be minimized to prevent contamination and cross-transmission. Point-of-care ultrasound, that is ultrasound performed and interpreted by a treating practitioner, has emerged as a potential solution to address early detection and monitoring.  Point-of-care lung ultrasound accurately detects acute respiratory distress syndrome (ARDS) when used by experts, and uses inexpensive, portable and easily cleaned hardware.  However, lung ultrasound can be challenging for novice users to perform and lung ultrasound images are difficult for non-expert users to interpret.
Picture ​Rapid automated analysis of lung ultrasound images by artificial intelligence (AI) comprises a diagnostic test minimally trained users could use to identify and monitor Covid19 patients with lung involvement. We are seeking to investigate this as a potential diagnostic and/or prognostic tool in patients with covid-19.

Purpose & hypothesis

To determine (1) whether it is feasible to obtain lung ultrasound imaging in Covid19 patients in emergency triage and intensive care unit (ICU) pandemic settings, and (2) the level of accuracy, reliability and prognostic information available from automated AI detection of Covid19 lung involvement.

We expect that (1) lung ultrasound imaging can be obtained efficiently by health care personnel in the covid19 pandemic setting, and (2) AI can detect covid19 lung involvement from lung ultrasound as accurately as human expert observers.

Study design

Study design: this is a multi-institution prospective study (UAH and RAH sites, ER and ICU settings).

Participants and setting: Consecutive consenting adult patients with a clinical history and/or laboratory test results consistent with active covid19 infection, presenting in two settings: (a) hospital emergency room (ER) for assessment, and (b) ICU admission due to respiratory decompensation. 

Methods

Methods: one portable ultrasound unit (e.g., Philips Lumify) and sterilization equipment (e.g., disposable sheaths and gel) will be provided for each study site.  Lung ultrasound experts on the study team will identify 3-5 interested staff routinely involved in direct care of Covid19 patients (e.g., nurses, respiratory therapists, residents, nurse practitioners or attending physicians) per study site and provide brief hands-on in-person training sessions as to how to obtain and label lung ultrasound images, focusing on lung bases. In both ER and ICU settings a baseline ultrasound of both lungs will be performed on each eligible patient by these trained study personnel.  In ICU setting, follow-up ultrasound will be obtained daily.  Any additional relevant imaging (CXR, chest CT) obtained for clinical care will be collected for correlative review. Images will be reviewed by team experts (intensivist and radiologist), initially blinded to clinical data and then combining all available clinical and imaging data, to determine (a) scan quality (0=inadequate for diagnosis or wrong body part, to 5=excellent diagnostic images), and (b) the presence and extent of lung abnormalities including (1) consolidation, (2) basal (pleural-parenchymal) interstitial abnormalities typical of ARDS, (3) interstitial abnormalities typical for pulmonary edema, and (4) effusion, by semiquantitative scoring system.  

Artificial intelligence: an AI network will be trained initially in lung ultrasound feature recognition using data available from a retrospective CXR/US study (n=300 patients), then refined for Covid19 using our prospective study data.  

​The AI network will be designed to predict (1) the probability of each imaging feature being present per-lung region and per-patient, and (adding clinical data to the images as network inputs) the probability of each clinical outcome (resolution, hospitalization, ICU admission, death).  

Will you be participating in this investigation through performing exams?

Please select this button to learn how to acquire clips for this study. Learn More

en_USEnglish