In Methods (San Diego, Calif.)
PURPOSE : In this paper, we utilized deep learning methods to screen the positive COVID-19 cases in chest CT. Our primary goal is to supply rapid and precise assistance for disease surveillance on the medical imaging aspect.
MATERIALS AND METHODS : Basing on deep learning, we combined semantic segmentation and object detection methods to study the lesion performance of COVID-19. We put forward a novel end-to-end model which takes advantage of the Spatio-temporal features. Furthermore, a segmentation model attached with a fully connected CRF was designed for a more effective ROI input.
RESULTS : Our method showed a better performance across different metrics against the comparison models. Moreover, our strategy highlighted strong robustness for the processed augmented testing samples.
CONCLUSION : The comprehensive fusion of Spatio-temporal correlations can exploit more valuable features for locating target regions, and this mechanism is friendly to detect tiny lesions. Although it remains in discrete form, the feature extracting in temporal dimension improves the precision of final prediction.
Jingxin Liu, Mengchao Zhang, Yuchen Liu, Jinglei Cui, Yutong Zhong, Zhong Zhang, Lihui Zu
COVID-19, Chest CT, Deep Learning, Object Detection, Semantic Segmentation