Communications in Information and Systems

Volume 20 (2020)

Number 4

CasCRNN-GL-Net: cascaded convolutional and recurrent neural networks with global and local pathways for classification of focal liver lesions in multi-phase CT images

Pages: 415 – 442

DOI: https://dx.doi.org/10.4310/CIS.2020.v20.n4.a2

Authors

Dong Liang (College of Computer Science and Technology, Zhejiang University, Hangzhou, China)

Yingying Xu (College of Computer Science and Technology, Zhejiang University, Hangzhou, China)

Lanfen Lin (College of Computer Science and Technology, Zhejiang University, Hangzhou, China)

Nan Zhou (College of Computer Science and Technology, Zhejiang University, Hangzhou, China)

Hongjie Hu (Department of Radiology, Sir Run Run Shaw Hospital, Hangzhou, China)

Qiaowei Zhang (Department of Radiology, Sir Run Run Shaw Hospital, Hangzhou, China)

Qingqing Chen (Department of Radiology, Sir Run Run Shaw Hospital, Hangzhou, China)

Xianhua Han (College of Information Science and Engineering, Ritsumeikan University, Kyoto, Japan)

Yutaro Iwamoto (College of Information Science and Engineering, Ritsumeikan University, Kyoto, Japan)

Yen-Wei Chen (College of Information Science and Engineering, Ritsumeikan University, Kyoto, Japan)

Abstract

Automatic focal liver lesion (FLL) classification based on multiphase computed tomography (CT) images is one of the most crucial components of a computer-aided diagnosis (CAD) system for liver disease. Though much research has been conducted in this field, two challenges still remain: (1) difficulty of representing the temporal enhancement pattern effectively and (2) the need to obtain the local and global information of FLLs. However, most of existing studies only focus on local or global information. In this paper, we propose a cascaded convolutional neural network (CNN) and recurrent neural network (RNN) with global and local pathways, called CasCRNN-GL-Net, for the classification of FLLs in multiphase CT images. The CNN with global and local pathways is used to extract both global and local information. A bi-directional long short-term memory (BD-LSTM) model is employed to represent the temporal enhancement pattern in multi-phase CT images. Moreover, we propose a novel joint loss function to train the proposed framework. The joint loss function is composed of interclass and intra-class losses, which can improve the robustness of the framework. The proposed method outperformed state-of-the-art approaches by achieving a mean accuracy of 87.64%. We have released the code on GitHub: $\href{https://github.com/UpCoder/GL_BD_LSTM}{\small{\texttt{https://github.com/UpCoder/GL_BD_LSTM}}}$.

Dong Liang and Yingying Xu contributed equally to this work.

The corresponding authors are Lanfen Lin, Hongjie Hu, and Yen-Wei Chen.

Received 3 August 2020

Published 2 December 2020