ACSN: Attention capsule sampling network for diagnosing COVID-19 based on chest CT scans

• Published in: Computers in biology and medicine Vol. 153; p. 106338
• Main Authors: Wen, Cuihong, Liu, Shaowu, Liu, Shuai, Heidari, Ali Asghar, Hijji, Mohammad, Zarco, Carmen, Muhammad, Khan
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.02.2023; Elsevier Limited
• Subjects: Accuracy; Algorithms; Capsule network; Chest; Chest CT scan; Classification; Computed tomography; Coronaviruses; COVID-19; COVID-19 - diagnostic imaging; COVID-19 recognition; COVID-19 Testing; Datasets; Deep learning; Disease; Disease control; Epidemics; Feature sampling; Humans; Lung infections; Medical diagnosis; Medical imaging; Medical treatment; Methods; Neural networks; Pandemics; Patients; Pneumonia; ROC Curve; Sampling; Sampling methods; SARS-CoV-2; Severe acute respiratory syndrome; Severe acute respiratory syndrome coronavirus 2; Thorax; Tomography, X-Ray Computed - methods; Viral diseases; Deep learning; COVID-19 recognition; Chest CT scan; Lung infections; Capsule network; Feature sampling
• ISSN: 0010-4825; 1879-0534; 1879-0534
• DOI: 10.1016/j.compbiomed.2022.106338

Automated diagnostic techniques based on computed tomography (CT) scans of the chest for the coronavirus disease (COVID-19) help physicians detect suspected cases rapidly and precisely, which is critical in providing timely medical treatment and preventing the spread of epidemic outbreaks. Existing capsule networks have played a significant role in automatic COVID-19 detection systems based on small datasets. However, extracting key slices is difficult because CT scans typically show many scattered lesion sections. In addition, existing max pooling sampling methods cannot effectively fuse the features from multiple regions. Therefore, in this study, we propose an attention capsule sampling network (ACSN) to detect COVID-19 based on chest CT scans. A key slices enhancement method is used to obtain critical information from a large number of slices by applying attention enhancement to key slices. Then, the lost active and background features are retained by integrating two types of sampling. The results of experiments on an open dataset of 35,000 slices show that the proposed ACSN achieve high performance compared with state-of-the-art models and exhibits 96.3% accuracy, 98.8% sensitivity, 93.8% specificity, and 98.3% area under the receiver operating characteristic curve. •Proposal of a key slices' enhancement method for accurate classification.•A key pooling sampling method for effective feature maps selection.•ACSN as a new network structure design for solving the problems of CT-Caps.