Improvement and Multi-Population Generalizability of a Deep Learning-Based Chest Radiograph Severity Score for COVID-19

摘要: ABSTRACT Purpose To improve and test the generalizability of a deep learning-based model for assessment COVID-19 lung disease severity on chest radiographs (CXRs) from different patient populations. Materials Methods A published convolutional Siamese neural network-based previously trained hospitalized patients with was tuned using 250 outpatient CXRs. This produces quantitative measure (pulmonary x-ray (PXS) score). The evaluated CXRs four sets, including 3 United States (patients at an academic medical center (N=154), community hospital (N=113), outpatients (N=108)) 1 Brazil emergency department (N=303)). Radiologists both countries independently assigned reference standard CXR scores, which were correlated PXS scores as performance (Pearson r). Uniform Manifold Approximation Projection (UMAP) technique used to visualize network results. Results Tuning learning data improved in two datasets (r=0.88 r=0.90, compared baseline r=0.86). Model similar, though slightly lower, when tested (r=0.86 r=0.85, respectively). UMAP showed that learned information generalized across sets. Conclusions Performance extracts score training cohort (outpatient versus hospitalized) multiple