Abstract
Reverse Transcription-Polymerase Chain Reaction (RT-PCR) is the gold standard for diagnosis of SARS-CoV-2 infection, but requires specialized equipment and reagents and suffers from long turnaround times. While valuable, chest imaging currently only detects COVID-19 pneumonia, but if it can predict actual RT-PCR SARS-CoV-2 status is unknown. Radiogenomics may provide an effective and accurate RT-PCR-based surrogate. We describe a deep learning radiogenomics (DLR) model (RadGen) that predicts a patient's RT-PCR SARS-CoV-2 status solely from their frontal chest radiograph (CXR).
Brief Introduction
Reverse Transcription-Polymerase Chain Reaction (RT-PCR) is the gold standard for diagnosis of SARS-CoV-2 infection1, but requires specialized equipment and reagents and suffers from long turnaround times. While valuable, chest imaging currently only detects COVID-19 pneumonia, but if it can predict actual RT-PCR SARS-CoV-2 status is unknown. Radiogenomics may provide an effective and accurate RT-PCR-based surrogate2. We describe a deep learning radiogenomics (DLR) model (RadGen) that predicts a patient’s RT-PCR SARS-CoV-2 status solely from their frontal chest radiograph (CXR).
Methods
The RadGen architecture, based on SE-ResNeXt-50-32×4d, was pretrained on ImageNet and ChestX-ray14 and 28,430 CXR from PadChest, and Kaggle before fine-tuned using CXR from a multinational cohort of RT-PCR tested patients from Hong Kong, GITHUB, SIRM and BIMCV (6,326 images)3-5. The model first predicted and selected only frontal CXR images, then predicted a segmentation mask of the cropped lung areas to reduce model fitting to unrelated parts of the image before using the segmented area as input for the RT-PCR SARS-CoV-2 binary classification task. The final prediction score was an ensemble consisting of the average of 4 models.
Results
RadGen achieved a mean Area Under the ROC curve (AUROC) of 0.959 (95%CI 0.955,0.962), sensitivity of 80.8% (3007/3723) and specificity 95.1% (16206/17033) using a pre-determined 0.4 cutpoint. It distinguishes pre-COVID-19, laboratory confirmed pneumonia from SARS2-CoV-2 cases with a specificity of 89.3% (225/252), and a specificity of 96.4% (106/110) on excluding SARS-CoV-2 infection on patients pre-COVID-19 CXR who later became RT-PCR positive.
The RT-PCR-tested cohort in Hong Kong consisted of 314 positive and 2,471 negative patients from 4 hospitals3. The SARS-CoV-2 positive patients had a median of 9 (range 3-24) RT-PCR tests and 3 corresponding (range 1-8) CXR during their hospitalization (range 3-112 days); 90.8% (285/314) had mild/asymptomatic disease. The sensitivity and specificity of RadGen for predicting SARS-CoV-2 infection on initial presenting CXR was 79.5% (225/283) and 85.2% (2105/2471).
RadGen time course analysis by autocorrelation function (ACF) plot, which describes how well RadGen predicts a patient’s RT-PCR SARS-CoV-2 status over the course of their entire SARS-CoV-2 infection period, was performed revealing a peak lag at 2 days for radiogenomic signature manifestation on CXR after initial RT-PCR diagnosis. The per-film false negative rate was 10.0% (26/261) and 9.3% (21/225) within and after 7 days of the first RT-PCR positive test with a false positive rate of 68.1% (32/47) and 11.3% (45/397) within and after one week of achieving RT-PCR confirmed viral clearance (Fig 1).
Serial CXR of a SARS-CoV-2 RT-PCR positive male patient in his 40s. RadGen correctly predicted the RT-PCR status of the patient on CXR prior to COVID19 (pre-COVID-19), at initial SARS-CoV-2 RT-PCR confirmed diagnosis (Day 0), during his infection period (Day 8) and upon achieving RT-PCR confirmed viral clearance (Day 12).
Comment
Leveraging a DLR strategy and a rich body of training datasets including Asian and Western countries (reflecting a diverse set of clinical containment protocols), a wide spectrum of clinical presentations including mild and asymptomatic disease, and a prospectively collected multi-timepoint RT-PCR SARS-CoV-2 positive patient cohort, we generated a DLR model capable of predicting a patient’s RT-PCR status from CXR.
Interestingly, we also show that RadGen can non-invasively ‘track’ RT-PCR SARS-CoV-2 status over the course of their infection, from diagnosis to viral clearance. A time-delayed correlation between RadGen and RT-PCR seen at the time of initial RT-PCR positivity and at the time of achieving RT-PCR viral clearance was observed. This is not unexpected as SARS-CoV-2 genomic dosage changes have been shown to take time to accumulate and be phenotypically reflected on a cellular, organ and systems level. Further, it is known that SARS-CoV-2 RNA can persist long after active infectivity and symptom resolution6; thus, that RadGen performs this well, particularly in a mild/asymptomatic cohort, is notable.
In conclusion, the feasibility for DLR models to dynamically track RT-PCR SARS-CoV-2 changes on an individual level significantly expands the scope of radiogenomics.
Data Availability
Code and models used in this study are available upon reasonable request to the corresponding author and under a collaboration agreement.
