COVID-19 diagnosis from chest X-ray images using transfer learning: Enhanced performance by debiasing dataloader.

摘要: BACKGROUND: Chest X-ray imaging has been proved as a powerful diagnostic method to detect and diagnose COVID-19 cases due its easy accessibility, lower cost rapid time. OBJECTIVE: This study aims improve efficacy of screening infected patients using chest images with the help developed deep convolutional neural network model (CNN) entitled nCoV-NET. METHODS: To train evaluate performance model, three datasets were collected from resources "ChestX-ray14", "COVID-19 image data collection", "Chest collection Indiana University," respectively. Overall, 299 pneumonia 1,522 non-COVID 19 are involved in this study. overcome probable bias unbalanced two classes datasets, ResNet, DenseNet, VGG architectures re-trained fine-tuning stage process distinguish transfer learning method. Lastly, optimized final nCoV-NET was applied testing dataset verify proposed model. RESULTS: Although parameters all determined close each other, nCOV-NET by DenseNet-161 architecture exhibits highest for classification accuracy 97.1 %. The Activation Mapping used create activation maps that highlights crucial areas radiograph causality intelligibility. CONCLUSION: demonstrated CNN called can be utilized reliably detecting accelerate triaging save critical time disease control well assisting radiologist validate their initial diagnosis.