2021 Volume Annual59 Issue Abstract Pages 311
We report on the development of KindAI-COVID, a system using deep-learning to differentiate COVID-19 pneumonia using chest CT images. Our system's feature is to differentiate COVID-19 pneumonia for each CT slice to collect many training images for deep-learning. First, the range of slices needed for diagnosis is specified by an experienced radiologist. Next, the lung fields are extracted using U-net. The images are differentiated using fine-tuned GoogLeNet pre-trained on the ImageNet. In this study, CT data of 39 patients with COVID-19 pneumonia and 91 patients with non-COVID-19 pneumonia taken at Kindai University Hospital are included. The accuracies were 83.1% and 87.2% per slice and per examination, respectively. Our system was able to differentiate COVID-19 pneumonia with relatively high accuracy. We believe that limiting the target to only the slices needed for diagnosis and training deep-learning on many slices were effective.
