Diagnosis Model of COVID-19 with Feature Loss

Abstract

As COVID-19 epidemic continues to spread globally, early detection and diagnosis are critical factors to control the virus spreading. In this paper, a COVID-19 intelligent diagnosis model from chest CT image is proposed, which can identify the positive CT and locate the lesions. The model first uses the U2-Net network to segment the lung parenchyma from the CT image, then the segmented image is input into the COVID-19 screening network based on EfficientNet to predict the result of infection with COVID-19 and lesion areas. A new loss function is proposed to constrain the network to extract effective lesion features. Grad-CAM algorithm is finally used to visualize the lesions in CT images. The dataset is collected from 1170 patients in Union Hospital (HUST-UH) and Liyuan Hospital (HUST-LH). The experimental results show that the dice coefficient of the segmentation network reaches 0.9838, the sensitivity of the COVID-19 screening network reaches 0.974, with an AUC of 0.994. The model trained with proposed loss function can locate the lesion areas more accurately.