Multimodal Glioma Image Segmentation Using Dual Encoder Structure and Channel Spatial Attention Block

Run Su; Jinhuai Liu; Deyun Zhang; Chuandong Cheng; Mingquan Ye

doi:10.3389/fnins.2020.586197

Multimodal Glioma Image Segmentation Using Dual Encoder Structure and Channel Spatial Attention Block

Front Neurosci. 2020 Oct 28:14:586197. doi: 10.3389/fnins.2020.586197. eCollection 2020.

Authors

Run Su^{1

2}, Jinhuai Liu^{1

2}, Deyun Zhang³, Chuandong Cheng^{4

5

6}, Mingquan Ye⁷

Affiliations

¹ Institute of Intelligent Machines, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, China.
² Science Island Branch of Graduate School, University of Science and Technology of China, Hefei, China.
³ School of Engineering, Anhui Agricultural University, Hefei, China.
⁴ Department of Neurosurgery, The First Affiliated Hospital of University of Science and Technology of China, Hefei, China.
⁵ Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, China.
⁶ Anhui Province Key Laboratory of Brain Function and Brain Disease, Hefei, China.
⁷ School of Medical Information, Wannan Medical College, Wuhu, China.

Abstract

Multimodal medical images provide significant amounts of complementary semantic information. Therefore, multimodal medical imaging has been widely used in the segmentation of gliomas through computational neural networks. However, inputting images from different sources directly to the network does not achieve the best segmentation effect. This paper describes a convolutional neural network called F-S-Net that fuses the information from multimodal medical images and uses the semantic information contained within these images for glioma segmentation. The architecture of F-S-Net is formed by cascading two sub-networks. The first sub-network projects the multimodal medical images into the same semantic space, which ensures they have the same semantic metric. The second sub-network uses a dual encoder structure (DES) and a channel spatial attention block (CSAB) to extract more detailed information and focus on the lesion area. DES and CSAB are integrated into U-Net architectures. A multimodal glioma dataset collected by Yijishan Hospital of Wannan Medical College is used to train and evaluate the network. F-S-Net is found to achieve a dice coefficient of 0.9052 and Jaccard similarity of 0.8280, outperforming several previous segmentation methods.

Keywords: CSAB; DES; F-S-Net; fully convolutional neural networks; glioma segmentation; medical image fusion.