Online COVID-19 diagnosis with chest CT images: Lesion-attention deep neural networks

Liu, B.; Gao, X.; He, M.; Lv, F.; Yin, G.

Chest (computed tomography) CT scanning is one of the most important technologies for COVID-19 diagnosis in the current clinical practice, which motivates more concerted efforts in developing AI-based diagnostic tools to alleviate the enormous burden on the medical system. We develop a lesion-attention deep neural network (LA-DNN) to predict COVID-19 positive or negative with a richly annotated chest CT image dataset. The CT image dataset contains 746 public chest CT images of COVID-19 patients collected from over 760 preprints, and the data annotations are accompanied with the textual radiology reports. We extract two types of important information from these annotations: One is the flag of whether an image indicates a positive or negative case of COVID-19, and the other is the description of five lesions on the CT images associated with the positive cases. The proposed data-driven LA-DNN model focuses on the primary task of binary classification for COVID-19 diagnosis, while an auxiliary multi-label learning task is implemented simultaneously to draw the model's attention to the five lesions of COVID-19 during the training. The joint task learning process makes it a highly sample-efficient deep model that can learn COVID-19 radiology features effectively with very limited samples. The experimental results show that the area under the curve (AUC) and sensitivity (recall) for the diagnosis of COVID-19 patients are 91.2% and 85.7% respectively, which reach the clinical standards for practical use. An online system has been developed for fast online diagnoses using CT images at the web address https://www.covidct.cn/.