Please click here to download the model.
The coronavirus disease 2019 (COVID-19) is raging inside China and internationally. We develop a deep learning model to discriminate patients with COVID-19 from other pneumonias using routine CT images of 807 patients (six centers). Patient from three of the centers were randomly divided into a training cohort (n=346) and an internal validation cohort (n=152). Patients from the other three centers were enrolled as an external validation cohort (n=309). We built a 10-layer 3D deep residual network as our model (termed as 3D-ResNet-10). The inputs were resized and cropped the lung regions to the size of (20, 243, 243). We trained the 3D-ResNet-10 on the training cohort using a NVIDIA Titan RTX Graphics Card for 60 epochs. The batch size was 64. We set a learning rate of 1e-5 and a weight-decay of 1. The binary cross entropy loss was adopted as the model’s criterion. The Adam optimizer was applied to step the model’s weights. Finally, the model achieves a impressive performance.