3D multi‐resolution attention capsule network for diagnosing multi‐pathological types of pulmonary nodules. Issue 5 (15th March 2022)

• Record Type: Journal Article
• Authors: Zhang, Zhili; Xiao, Taohui; Fu, Yu; Gao, Yuqiang; Ren, Meirong; Cui, Wentao; Dong, Enqing
• Is Part Of: International journal of imaging systems and technology. Volume 32:Issue 5(2022)
• Page Start: 1727
• Page End: 1742
• Publication Date: 2022-03-15
• Subjects: attention mechanism -- capsule network -- deep learning -- multi‐pathological classification -- pulmonary nodule diagnosis; Imaging systems -- Periodicals; Image processing -- Periodicals; 621.367
• DOI: 10.1002/ima.22726
• Languages: English
• ISSNs: 0899-9457

Abstract

Abstract: To achieve effective diagnosis of multi‐pathological types on pulmonary nodules, a three‐dimensional (3D) multi‐resolution attention (MRA) capsule network (3D MRA‐CapsNet‐DL) was proposed by using a self‐constructed lung CT image dataset. In the construction of 3D MRA‐CapsNet‐DL, an improved dynamic routing algorithm (DRA) with limiting the update increment (called DL) was first designed to solve the inactivation problem of a large number of vector neurons. Then, MRA was used to build three kinds of multi‐path 3D MRA‐CapsNet‐DL for multi‐pathological classification. Besides, we also investigated whether the interpolation perturbation introduced by multi‐resolution methods will affect the classification performance. Experiments showed that the proposed 3D MRA‐CapsNet‐DL can achieve a better classification (79.27% ACC and 0.9056 AUC) effect than classical convolutional neural networks (CNNs), and the interpolation disturbance has favorable and unfavorable effects on the classification performance of multi‐path and single‐path models, respectively.