Accurate detection of COVID-19 patterns from CT lung scans by deep learning
In our laboratory, we have been extensively working with the exact same technology and have applied it to similar CT images to detect areas of epicardial fat around the heart, and calcium, and applied it also to nuclear cardiology images. Importantly, these models are already clinically deployed for testing. Our publication with deep learning for epicardial fat segmentation from CT was highlighted in Radiology: Artificial Intelligence, deployed in QFAT software and the prognostic value of automated AI-based epicardial fat has been published for the prospective EISNER trial at Cedars-Sinai. Several of our publications in nuclear cardiology show great improvement over visual reading with the deep learning approach. We, therefore, have immediate capabilities in our lab at Cedars-Sinai to develop similar or improved AI technology, to detect specific COVID-19 lung damage on CT. CT data with such COVID-19 patterns would be essential to conduct this project.
We propose a rapid and multipronged approach to develop state-of-the art deep learning detection of COVID-19 damage, leveraging our extensive experience in deep learning and CT image processing. We are actively seeking data from collaborators internationally and will develop an internal database of confirmed COVID-19 CT cases and controls at Cedars Sinai. We will utilize the NLST dataset as controls to differentiate between the COVID patterns found in our center and collaborating centers. The use of NLST data set for this purpose will allow for superior specificity for detection of COVID-19 changes
We will divert our current technical staff currently working on deep learning calcium detection from chest CT to adapt these algorithms for the accurate detection of COVID-19 patterns in the lungs. All the infrastructure is in place in our laboratory, including necessary hardware with graphics processing units (GPU), and data anonymization tools. All CT data will be anonymized at source. Waiver for informed consent will be requested to analyze de-identified data.
We will utilize 3D deep convolutional network (CNN) framework (U-net and/or DenseNet) approaches, explainable AI-tools, GradCAM approach to visualize to the physicians which specific region of the lungs indicate COVID-19 infection. All analysis will be compared with interobserver visual analysis and RT-PCR confirmation for accuracy. Appropriate thresholds for optimal sensitivity and specificity will be developed. Tools for monitoring of progression of the disease from serial scans will be also developed. We except that doctors assisted with AI tools will perform much faster, more accurately and with higher confidence, and with lower interobserver variabilities in the detection of specific COVID-19 patterns. The prediction will be evaluated in independent hold-out sets.
1) Develop deep learning system for detection of COVD-19 patterns from chest CT
2) validate the system sensitivity with the RT-PCT positive CT chest scans from Cedars-Sinai center
3) Validate the system specificity in the NLST chest CT dataset
Damini Dey, Cedars Sinai
Barry Pressman, Cedars-Sinai
Peter Julien, Cedars -Sinai