부산대학교 도서관

Introduction: Epidemiological studies that involve interpretation of chest radiographs (CXRs) suffer from inter-reader and intra-reader variability. Inter-reader and intra-reader variability hinder comparison of results from different studies or centres, which negatively affects efforts to track the burden of chest diseases or evaluate the efficacy of interventions such as vaccines. This study explores machine learning models that could standardize interpretation of CXR across studies and the utility of incorporating individual reader annotations when training models using CXR data sets annotated by multiple readers. Methods: Convolutional neural networks were used to classify CXRs from seven low to middle-income countries into five categories according to the World Health Organization's standardized methodology for interpreting paediatric CXRs. We compared models trained to predict the final/aggregate classification with models trained to predict how each reader would classify an image and then aggregate predictions for all readers using unweighted mean. Results: Incorporating individual reader's annotations during model training improved classification accuracy by 3.4% (multi-class accuracy 61% vs 59%). Model accuracy was higher for children above 12 months of age (68% vs 58%). The accuracy of the models in different countries ranged between 45% and 71%. Conclusions: Machine learning models can annotate CXRs in epidemiological studies reducing inter-reader and intra-reader variability. In addition, incorporating individual reader annotations can improve the performance of machine learning models trained using CXRs annotated by multiple readers.