MHA-CoroCapsule: Multi-Head Attention Routing-Based Capsule Network for COVID-19 Chest X-Ray Image Classification

• Published in: IEEE transactions on medical imaging Vol. 41; no. 5; pp. 1208 - 1218
• Main Authors: Li, Fudong, Lu, Xingyu, Yuan, Jianjun
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.05.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Annotations; Attention; capsule networks; Chest; chest X-ray images; Convolution; Coronaviruses; COVID-19; COVID-19 - diagnostic imaging; Datasets; Deep Learning; Diagnosis; Feature extraction; Health care; Humans; Image classification; Iterative methods; Machine learning; Medical imaging; Model accuracy; multi-head attention; Neural Networks, Computer; Physicians; Pulmonary diseases; Routing; SARS-CoV-2; Transfer learning; Viral diseases; X-ray imaging; X-Rays
• ISSN: 0278-0062; 1558-254X; 1558-254X
• DOI: 10.1109/TMI.2021.3134270

The outbreak of COVID-19 threatens the lives and property safety of countless people and brings a tremendous pressure to health care systems worldwide. The principal challenge in the fight against this disease is the lack of efficient detection methods. AI-assisted diagnosis based on deep learning can detect COVID-19 cases for chest X-ray images automatically, and also improve the accuracy and efficiency of doctors' diagnosis. However, large scale annotation of chest X-ray images is difficult because of limited resources and heavy burden on the medical system. To meet the challenge, we propose a capsule network model with multi-head attention routing algorithm, called MHA-CoroCapsule, to provide fast and accurate diagnostics for COVID-19 diseases from chest X-ray images. The MHA-CoroCapsule consists of convolutional layers, two capsule layers, and a non-iterative, parameterized multi-head attention routing algorithm is used to quantify the relationship between the two capsule layers. The experiments are performed on a combined dataset constituted by two publicly available datasets including normal, non-COVID pneumonia and COVID-19 images. The model achieves the accuracy of 97.28%, recall of 97.36%, and precision of 97.38% even with a limited number of samples. The experimental results demonstrate that, contrary to the transfer learning and deep feature extraction approaches, the proposed MHA-CoroCapsule has an encouraging performance with fewer trainable parameters and does not require pretraining and plenty of training samples.