Premium
COVID‐19 diagnosis on CT scan images using a generative adversarial network and concatenated feature pyramid network with an attention mechanism
Author(s) -
Li Zonggui,
Zhang Junhua,
Li Bo,
Gu Xiaoying,
Luo Xudong
Publication year - 2021
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1002/mp.15044
Subject(s) - artificial intelligence , computer science , pattern recognition (psychology) , covid-19 , feature (linguistics) , pyramid (geometry) , generative adversarial network , precision and recall , classifier (uml) , computed tomography , deep learning , mathematics , radiology , medicine , infectious disease (medical specialty) , pathology , linguistics , philosophy , geometry , disease
Objective Coronavirus disease 2019 (COVID‐19) has caused hundreds of thousands of infections and deaths. Efficient diagnostic methods could help curb its global spread. The purpose of this study was to develop and evaluate a method for accurately diagnosing COVID‐19 based on computed tomography (CT) scans in real time. Methods We propose an architecture named “concatenated feature pyramid network” (“Concat‐FPN”) with an attention mechanism, by concatenating feature maps of multiple. The proposed architecture is then used to form two networks, which we call COVID‐CT‐GAN and COVID‐CT‐DenseNet, the former for data augmentation and the latter for data classification. Results The proposed method is evaluated on 3 different numbers of magnitude of COVID‐19 CT datasets. Compared with the method without GANs for data augmentation or the original network auxiliary classifier generative adversarial network, COVID‐CT‐GAN increases the accuracy by 2% to 3%, the recall by 2% to 4%, the precision by 1% to 3%, the F1‐score by 1% to 3%, and the area under the curve by 1% to 4%. Compared with the original network DenseNet‐201, COVID‐CT‐DenseNet increases the accuracy by 1% to 3%, the recall by 4% to 9%, the precision by 1%, the F1‐score by 1% to 3%, and the area under the curve by 2%. Conclusion The experimental results show that our method improves the efficiency of diagnosing COVID‐19 on CT images, and helps overcome the problem of limited training data when using deep learning methods to diagnose COVID‐19. Significance Our method can help clinicians build deep learning models using their private datasets to achieve automatic diagnosis of COVID‐19 with a high precision.