Journal of Computer Applications ›› 2021, Vol. 41 ›› Issue (6): 1820-1827.DOI: 10.11772/j.issn.1001-9081.2020111788

Special Issue: 前沿与综合应用

• Frontier and comprehensive applications • Previous Articles     Next Articles

Segmentation of ischemic stroke lesion based on long-distance dependency encoding and deep residual U-Net

HUANG Li, LU Long   

  1. School of Information Management, Wuhan University, Wuhan Hubei 430072, China
  • Received:2020-11-16 Revised:2021-01-08 Online:2021-06-10 Published:2021-05-08
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61772375), the National Key Research and Development Program of China (2019YFC0120000), the Hubei Provincial Natural Science Foundation (2019CFA025), the Independent Research Project of School of Information Management in Wuhan University (413100032).

基于长距离依赖编码与深度残差U-Net的缺血性卒中病灶分割

黄梨, 卢龙   

  1. 武汉大学 信息管理学院, 武汉430072
  • 通讯作者: 卢龙
  • 作者简介:黄梨(1996-),女,湖南衡阳人,硕士研究生,主要研究方向:医学影像分析、数据挖掘、深度学习;卢龙(1976-),男,福建福清人,教授,博士,主要研究方向:医学影像分析、生物信息学、深度学习。
  • 基金资助:
    国家自然科学基金资助项目(61772375);国家重点研发计划项目(2019YFC0120000);湖北省自然科学基金资助项目(2019CFA025);武汉大学信息管理学院自主科研项目(413100032)。

Abstract: Segmenting stroke lesions automatically can provide valuable support to the clinical decision process. However, this is a challenging task due to the diversity of lesion size, shape, and location. Previous works have failed to capture global context information which is helpful to handle the diversity. To solve the problem of segmentation of ischemic stroke lesions with small sample size, an end-to-end neural network combing with residual block and non-local block on the basis of traditional U-Net was proposed to predict stroke lesion from multi-modal Magnetic Resonance Imaging (MRI) image. In this method, based on the encoder-decoder architecture of U-Net, residual blocks were stacked to solve the degradation problem and avoid the overfitting, and the non-local blocks were added to effectively encode the long-distance dependencies and provide global context information for the feature extraction process. The proposed method and its variants were evaluated on the Ischemic Stroke Lesion Segmentation (ISLES) 2017 dataset. The results showed that the proposed residual U-Net (Dice=0.29±0.23, ASSD=7.66±6.41, HD=43.71±22.11) and Residual Non-local U-Net (RN-UNet) (Dice=0.29±0.23, ASSD=7.61±6.62, HD=45.36±24.75) achieved significant improvement in all metrics compared to the baseline U-Net (Dice=0.25±0.23, ASSD=9.45±7.36, HD=54.59 ±21.19); compared with the state-of-the-art methods from ISLES website, the two methods both achieved better segmentation results, so that they can help doctors to quickly and objectively evaluate the condition of patients in clinical practices.

Key words: Acute Ischemic Stroke (AIS), residual learning, U-Net, medical image segmentation, Magnetic Resonance Imaging (MRI), deep learning

摘要: 脑卒中病灶自动分割可以为临床决策过程提供有价值的支持,而由于病灶大小、形状和位置的多样性,这项任务具有一定的挑战性。以往的研究未能很好地捕获有助于处理这种多样性的全局上下文信息。针对小样本情境下的缺血性脑卒中病灶分割这一问题,提出了在传统U-Net的基础上融合了残差模块和non-local块的端到端神经网络,用于从多模态核磁共振成像(MRI)的图像中预测卒中病灶。该方法基于U-Net的编码器-解码器结构,利用残差模块的堆叠来解决网络退化问题和避免过拟合,并通过插入non-local块编码特征图中的长距离依赖来为特征提取的过程提供全局上下文信息。对所提出的方法及其变体在缺血性卒中病灶分割挑战赛(ISLES)2017数据集上进行了评估,结果显示,所提出的残差U-Net (Dice=0.29±0.23、ASSD=7.66±6.41、HD=43.71±22.11)和RN-UNet (Dice=0.29±0.23、ASSD=7.61±6.62、HD=45.36±24.75),相对于基线U-Net (Dice=0.25±0.23、ASSD=9.45±7.36、HD=54.59±21.19)在所有指标上都有明显提升;跟ISLES官网上最先进的方法对比,所提的两个方法均取得了更好的分割结果,可见这两个方法有助于医生在临床实践中快速客观地评估病情。

关键词: 急性缺血性脑卒中, 残差学习, U-Net, 医学图像分割, 核磁共振成像, 深度学习

CLC Number: