ABSTRACT
COVID-19 associated lung diseases can mimic radiological characteristics of other viral lung diseases such as influenza which may lead to misdiagnosis. In this study, we proposed an Artificial Intelligence framework based on a combination of a Convolutional Neural network architecture and a Recurrent Neural Network architecture to classify CT volumes with COVID-19, Influenza, and no-infection. The model was trained on a dataset of 300 patients (100 patients in each class). Each set of 15 consecutive axial slices with the associated label of the corresponding CT volume was input as a 3 channel input at 5 time points to the CNN-RNN network. Benchmarked against RT-PCR confirmed cases of COVID-19 and Influenza, our model, when evaluated on an independent validation set of 400 CT patients, can accurately classify CT volumes of patients with COVID-19, Influenza, or no-infection with a sensitivity of 96% (COVID-19) and 95% (Influenza) (Tablel). Figurel shows the percentage of correctly classified and misclassified cases in each class. Our model provides rapid accurate diagnosis in patients suspected of COVID-19 infection, facilitating the timely implementation of isolation procedures and early intervention.
ABSTRACT
The coronavirus disease 2019 (COVID-19) outbreak has reached pandemic status and pushed healthcare systems beyond the limits. We aim to develop a fully automatic framework to detect COVID-19 by applying artificial intelligence (Al). A fully automated Al framework was developed to extract radiomics features from chest CT scans to detect COVID-19 patients. We curated and analysed the data from a total of 1381 patients. A cohort of 181 RT-PCR confirmed COVID-19 patients and 1200 control patients was included for model development. An independent dataset of 697 patients was used to validate the model. The datasets were collected from CHU Liège, Belgium. Model performance was assessed by the area under the receiver operating characteristic curve (AUC). Assuming 15% disease prevalence, a comprehensive analysis of classification performance in terms of accuracy, sensitivity, specificity, negative predictive value (NPV) and positive predictive value (PPV) was performed for all possible decision thresholds. The final curated dataset used for model development and testing consisted of chest CT scans of 1224 patients and 641 patients, respectively. The model had an AUC of 0.882 (95% CI: 0.851-0.913) in the independent test dataset. Assuming the cost of false negatives is twice as high as the cost of false positives, the optimal decision threshold resulted in an accuracy of 85.18%, a sensitivity of 69.52, a specificity of 91.63%, an NPV of 94.46% and a PPV of 59.44%. Our Al framework can accurately detect COVID-19. Thus, providing rapid accurate diagnosis in patients suspected of COVID-19 infection, facilitating the implementation of isolation procedures and early intervention.
ABSTRACT
In the course of the pandemic induced by the appearance of a new coronavirus (SARS-CoV-2;COVID-19) causing acute respiratory distress syndrome (ARDS), we had to rethink the diagnostic approach for patients suffering from respiratory symptoms. Indeed, although the use of RT-PCR remains the keystone of the diagnosis, the delay in diagnosis as well as the overload of the microbiological platforms have led us to make almost systematic the use of thoracic imaging for taking in charge of patients. In this context, thoracic imaging has shown a major interest in diagnostic aid in order to better guide the management of patients admitted to hospital. The most common signs encountered are particularly well described in thoracic computed tomography. Typical imaging combines bilateral, predominantly peripheral and posterior, multi-lobar, ground glass opacities. Of note, it is common to identify significant lesions in asymptomatic patients, with imaging sometimes preceding the onset of symptoms. Beyond conventional chest imaging, many teams have developed new artificial intelligence tools to better help clinicians in decision-making.