Performance of a computer aided diagnosis system for SARS-CoV-2 pneumonia based on ultrasound images.

ABSTRACT

PURPOSE: In this study we aimed to leverage deep learning to develop a computer aided diagnosis (CAD) system toward helping radiologists in the diagnosis of SARS-CoV-2 virus syndrome on Lung ultrasonography (LUS). METHOD: A CAD system is developed based on a transfer learning of a residual network (ResNet) to extract features on LUS and help radiologists to distinguish SARS-CoV-2 virus syndrome from healthy and non-SARS-CoV-2 pneumonia. A publicly available LUS dataset for SARS-CoV-2 virus syndrome consisting of 3909 images has been employed. Six radiologists with different experiences participated in the experiment. A comprehensive LUS data set was constructed and employed to train and verify the proposed method. Several metrics such as accuracy, recall, precision, and F1-score, are used to evaluate the performance of the proposed CAD approach. The performances of the radiologists with and without the help of CAD are also evaluated quantitively. The p-values of the t-test shows that with the help of the CAD system, both junior and senior radiologists significantly improve their diagnosis performance on both balanced and unbalanced datasets. RESULTS: Experimental results indicate the proposed CAD approach and the machine features from it can significantly improve the radiologists' performance in the SARS-CoV-2 virus syndrome diagnosis. With the help of the proposed CAD system, the junior and senior radiologists achieved F1-score values of 91.33% and 95.79% on balanced dataset and 94.20% and 96.43% on unbalanced dataset. The proposed approach is verified on an independent test dataset and reports promising performance. CONCLUSIONS: The proposed CAD system reports promising performance in facilitating radiologists' diagnosis SARS-CoV-2 virus syndrome and might assist the development of a fast, accessible screening method for pulmonary diseases.