«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年12期

页码:

25-32

栏目:

媒体计算

出版日期:

2024-12-10

文章信息/Info

Title:

Segmentation of Brain White Matter Lesions Using Fusion of Self-calibrated Convolution and Attention

文章编号:

1673-629X(2024)12-0025-08

作者:

岳江1; 刘庆晨1; 韩晓鑫1; 刘浩2; 王建林2*

1. 甘肃中医药大学 医学信息工程学院,甘肃 兰州 730000;2. 兰州大学第一医院,甘肃 兰州 730000

Author(s):

YUE Jiang1; LIU Qing-chen1; HAN Xiao-xin1; LIU Hao2; WANG Jian-lin2*

1. School of Medical Information Engineering,Gansu University of Traditional Chinese Medicine,Lanzhou 730000,China;2. Lanzhou University First Hospital,Lanzhou 730000,China

关键词:

脑白质病变; 分割; 脑白质高信号; 注意力机制; 自校正卷积

Keywords:

white matter lesions; segmentation; white matter hyperintensities; attention mechanism; SCConv

分类号:

TP391;R318

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0264

摘要:

脑白质病变是导致老年人认知功能障碍的主要原因之一,被认为是脑血管疾病的征兆。 原始 U-Net 模型在图像分割任务中面临的主要问题包括细小病变区域的漏检、边界分割不准确等。 提出一种基于自校正卷积与注意力机制的 2D U-Net 模型用于脑白质病变分割。 首先,引入自校正卷积模块,整合其周边区域的信息以及通道间的相互作用,提高对细微病变检测的准确性。 其次,使用两种不同的注意力模块,在编码的浅层和深层分别引入通道注意力机制和空间注意力机制,浅层编码器捕捉脑白质病变纹理的细粒度特征,而深层编码器提取病变的高级全局语义特征。 最后,采用了一种跨层融合策略,将解码器模块中的特征图通过 Transpose 操作与编码器同层的特征图进行尺度特征的整合。 实验结果表明,在 2017 WMH 分割挑战赛数据集和武汉同济医院数据集上分别测试了模型,其中 Dice 系数和 Recall 都分别达到了0. 80、0. 82 和 0. 82、0. 86。 该方法可以有效地检测出脑白质病变,并且在 1. 5T 磁共振成像协议识别效果显著。

Abstract:

White matter lesions are one of the main causes of cognitive dysfunction in the elderly and are considered to be signs of cere-brovascular disease. The main problems faced by the original U-Net model in image segmentation tasks include missed detection of small lesion areas and inaccurate boundary segmentation. A 2D U-Net model based on self-calibrated convolution and attention mechanism is proposed for white matter lesion segmentation. Firstly,a self-calibrated convolution module is introduced to integrate the information of its surrounding areas and the interaction between channels to improve the accuracy of subtle lesion detection. Secondly,two different attention modules are used to introduce channel attention mechanism and spatial attention mechanism in the shallow and deep layers of en-coding respectively. The shallow encoder captures the fine-grained features of the white matter lesion texture,while deep encoders extract high-level global semantic features of lesions. Finally,a cross-layer fusion strategy is adopted to integrate the scale features of the feature map in the decoder module with the feature map of the same layer of the encoder through the Transpose operation.Experimental results show that the model was tested on the 2017 WMH Segmentation Challenge data set and Wuhan Tongji Hospital data set,in which the Dice coefficient and Recall reached 0. 80,0. 82,and 0. 82,0. 86 respectively. The proposed method can effectively detect brain white matter lesions,and the recognition effect is remarkable in the 1. 5T magnetic resonance imaging protocol.

常用功能

更新日期/Last Update: 2024-12-10