Predicting COVID-19 pneumonia severity on chest X-ray with convolutional neural network: A retrospective study

OBJECTIVES: Radiological lung changes in COVID-19 infections present a noteworthy avenue to develop chest X-ray (CXR) -based testing models to support existing rapid detection techniques. The purpose of this study is to evaluate the accuracy of artificial intelligence (AI) -based screening model employing deep convolutional neural network for lung involvement. MATERIAL AND METHODS: An AI-based screening model was developed with state-of-the-art neural networks using Indian data sets from COVID-19 positive patients by authors of CAIR, DRDO, in collaboration with the other authors. Our dataset was comprised of 1324 COVID-19, 1108 Normal, and 1344 Pneumonia CXR images. Transfer learning was carried out on Indian dataset using popular deep neural networks, which includes DenseNet, ResNet50, and ResNet18 network architectures to classify CXRs into three categories. The model was retrospectively used to test CXRs from reverse transcriptase-polymerase chain reaction (RT-PCR) proven COVID-19 patients to test positive predictive value and accuracy. RESULTS: A total of 460 RT-PCR positive hospitalized patients CXRs in various stages of disease involvement were retrospectively analyzed. There were 248 males (53.92%) and 212 females (46.08%) in the cohort, with a mean age of 50.1 years (range 12–89 years). The commonly observed alterations included lung consolidations, ground-glass opacities, and reticular–nodular opacities. Bilateral involvement was more common compared to unilateral involvement. Of the 460 CXRs analyzed, the model reported 445 CXRs as COVID -19 with an accuracy of 96.73%. CONCLUSION: Our model, based on a two-level classification decision fusion and output information computation, makes it a robust, accurate and reproducible tool. Based on the initial promising results, our application can be used for mass screening.