White Matter Hyperintensities Segmentation Using Neural Network Ensembles

doi:10.3969/j.issn.1673-5765.2020.03.003

Abstract

Abstract:

Objective To construct a model applying the full convolutional neural network to improve the segmentation accuracy of white matter hyperintensities. Methods This model, which is based on the encoder-decoder architecture of U-Net, employs dense connections, and optimizes the number of convolutional layers to make full use of the features, and can implement end-to-end segmentation. Neural network ensembles are used to improve the robustness of the model. The performance metrics assessing this model included Dice similarity coefficient, Hausdorff distance, average volume difference, and F1-score. Results The experimental test results showed that this model achieved a high performance on various metrics (including Dice similarity coefficient, Hausdorff distance, average volume difference, and F1- score), outperforming the mainstream segmentation methods. The results proved the effectiveness of this method. Conclusions The models based on dense connections and ensemble can obtain a better model for white matter hyperintensities segmentation. This method provides an important algorithm reference for further analysis of white matter characteristics in cerebrovascular disease.

Key words: White matter hyperintensities segmentation; Deep learning; Dense connectivity;Ensemble model

摘要：

目的构建一种新型基于全卷积神经网络的分割方法，以提升脑白质高信号的分割精度。方法本模型基于U-Net的编码器-解码器结构，采用密集连接结构，并优化卷积次数，以充分利用网络中间层提取的特征，实现端到端的分割。该模型采用多个网络的集成框架，以提高模型的鲁棒性。模型对白质高信号的分割精度评价指标包括相似性系数、豪斯多夫距离、平均体积差异和F1分数。结果在公开数据集上进行的测试结果表明，本文提出的方法在4种分割评价标准（包括相似性系数，豪斯多夫距离，平均体积差异和F1分数）上的表现，优于现有的主流的分割方法，证明了该方法的有效性。结论基于密集连接和集成优化的神经网络模型，能对脑白质高信号进行较好的分割。该方法的提出，为进一步分析脑血管病白质特征，提供了重要的算法支撑。

关键词: 脑白质高信号分割; 深度学习; 密集连接; 多网络集成

LI Xin-Xin, WANG Xu-Xian,CHENG Jian, XU Hong, LI Zi-Xiao, LIU Tao. White Matter Hyperintensities Segmentation Using Neural Network Ensembles[J]. Chinese Journal of Stroke, 2020, 15(03): 234-242.

李鑫鑫*，汪绪先*，程健，徐红，李子孝，刘涛（*第一作者）. 基于多网络集成的脑白质高信号分割方法[J]. 中国卒中杂志, 2020, 15(03): 234-242.

References

[1] 中华医学会神经病学分会，中华医学会神经病学分

会脑血管病学组. 中国脑小血管病诊治共识[J]. 中华

神经科杂志，2015，48（10）：838-844.

[2] 楼敏，严慎强. 重视脑白质高信号的异质性和临床重

要性[J]. 中华医学杂志，2012，92（31）：2165-2167.

[3] WARDLAW J M，SMITH C，DICHGANS M.

Mechanisms of sporadic cerebral small vessel disease：

insights from neuroimaging[J]. Lancet Neurol，2013，

12（5）：483-497.

[4] SMITH E E，SCHNEIDER J A，WARDLAW J M，

et al. Cerebral microinfarcts：the invisible lesions[J].

Lancet Neurol，2012，11（3）：272-282.

[5] DEBETTE S，MARKUS H S. The clinical

importance of white matter hyperintensities on brain

magnetic resonance imaging：systematic review and

meta-analysis[J/OL]. BMJ，2010，341：c3666[2020-

01-10]. https：//doi.org/10.1136/bmj.c3666.

[6] D'ARBELOFF T，ELLIOTT M L，KNODT

A R，et al. White matter hyperintensities are

common in midlife and already associated with

cognitive decline[J/OL]. Brain Commun，2019，1

（1）：fcz041[2020-01-10]. https：//doi.org/10.1093/

braincomms/fcz041.

[7] HERRMANN L L，Le MASURIER M，EBMEIER

K P. White matter hyperintensities in late life

depression：a systematic review[J]. J Neurol

Neurosurg Psychiatry，2008，79（6）：619-624.

[8] YOSHITA M，FLETCHER E，HARVEY D，et

al. Extent and distribution of white matter

hyperintensities in normal aging，MCI，and AD[J].

Neurology，2006，67（12）：2192-2198.

[9] PHUAH C L，CHEN Y S，LIU Z Y，et al. White

matter hyperintensity spatial pattern variations

reflect distinct cerebral small vessel disease

pathologies[J/OL]. Stroke，2019，50（Suppl 1）：

A49[2020-01-10]. https：//doi.org/10.1161/str.50.

suppl_1.49.

[10] GRAFF-RADFORD J，de ARENAZA-URQUIJO

E，SCHWARZ C，et al. Topographic white matter

hyperintensity patterns associated with Alzheimer

pathologies[J/OL]. Stroke，2019，50（Suppl 1）：

A104[2020-01-10]. https：//doi.org/10.1161/str.50.

suppl_1.104.

[11] SÖDERLUND H，NILSSON L G，BERGER K，et

al. Cerebral changes on MRI and cognitive function：

the CASCADE study[J]. Neurobiol Aging，2006，27

（1）：16-23.

[12] KIM S H，PARK J S，AHN H J，et al. Voxel‐based

analysis of diffusion tensor imaging in patients with subcortical vascular cognitive impairment：

correlates with cognitive and motor deficits[J]. J

Neuroimaging，2011，21（4）：317-324.

[13] KUIJF H J，CASAMITJANA A，COLLINS D

L，et al. Standardized assessment of automatic

segmentation of white matter hyperintensities and

results of the WMH Segmentation Challenge[J].

IEEE Trans Med Imaging，2019，38（11）：2556-

2568.

[14] SHORTEN C，KHOSHGOFTAAR，T M. A survey

on image data augmentation for deep learning[J].

Journal of Big Data，2019，6（1）：60.

[15] RONNEBERGER O，FISCHER P，BROX T. U-net：

Convoluti onal networks for biomedical image

segmentation[C/OL]//International Conference on

Medical Image Computing and Computer-assisted

Intervention（MICCAI）2015. Springer，2015：234-

241[2020-01-10]. https：//doi.org/10.1007/978-3-319-

24574-4_28.

[16] PENG C，ZHANG X，YU G，et al. Large Kernel

Matters--Improve Semantic Segmentation by Global

Convolutional Network[C/OL]//Proceedings of the

IEEE Conference on Computer Vision and Pattern

Recognition 2017. 2017：4353-4361[2020-01-10].

https：//doi.org/10.1109/CVPR.2017.189.

[17] HUANG G，LIU Z，van der MAATEN L，et al.

Densely Connected Convolutional Networks[C/

OL]//2017 IEEE Conference on Computer Vision

and Pattern Recognition（CVPR）. IEEE，2017：

2261-2269[2020-01-10]. https：//doi.org/10.1109/

CVPR.2017.243.

[18] MILLETARI F，NAVAB N，AHMADI S A. V-Net：

Fully convolutional neural networks for volumetric

medical image segmentation[C/OL]//3D Vision

（3DV），2016 The 4th IEEE International conference

on Smart Energy Grid Engineering（SEGE 2016）.

IEEE，2016：565-571[2020-01-10]. https：//doi.

org/10.1109/3DV.2016.79.

[19] OPITZ D，MACLIN R. Popular ensemble methods：

an empirical study[J/OL]. J Artif Intell Res，1999，

11：169-198[2020-01-10]. https：//doi.org/10.1613/jair.

614.

[20] DIETTERICH T G. Ensemble Methods in Machine

Learning[C/OL]//International Workshop on

Multiple Classifier Systems（MCS）2000. Springer，

2000：1-15[2020-01-10]. https：//doi.org/10.1007/3-

540-45014-9_1.

[21] ZHOU Z H，WU J，TANG W. Ensembling neural

networks：many could be better than all[J]. Artificial

Intelligence，2002，137（1-2）：239-263.

[22] KAMNITSAS K，BAI W，FERRANTE E，et al.

Ensembles of multiple models and architectures

for robust brain tumour segmentation[C/OL]//

International MICCAI Brainlesion Workshop.

Springer，2017：450-462[2020-01-10]. https：//doi.

org/10.1007/978-3-319-75238-9_38.

[23] KITTLER J，HATER M，DUIN R P W. On

combining classifiers[C/OL]//Proceedings of 13th

International Conference On Pattern Recognition.

IEEE Transactions on Pattern Analysis and Machine

Intelligence，1996，20（3）：897-901[2020-01-10].

https：//doi.org/10.1109/ICPR.1996.547205.

[24] CHOLLET F. Keras[DB/OL].[2020-01-10]. https：//

github.com/fchollet/keras.

[25] LI H，JIANG G，ZHANG J，et al. Fully

convolutional network ensembles for white matter

hyperintensities segmentation in MR images[J/OL].

NeuroImage，2018，183：650-665[2020-01-10].

https：//doi.org/10.1016/j.neuroimage.2018.07.005.

[26] JIANG J，LIU T，ZHU W，et al. UBO Detector--

A cluster-based，fully automated pipeline for

extracting white matter hyperintensities[J/OL].

Neuroimage，2018，174：539-549[2020-01-10]. https：

//doi.org/10.1016/j.neuroimage.2018.03.050.

[27] DADAR M，MARANZANO J，MISQUITTA

K，et al. Performance comparison of 10 different

classification techniques in segmenting white matter

hyperintensities in aging[J/OL]. Neuroimage，2017，

157：233-249[2020-01-10]. https：//doi.org/10.1016/

j.neuroimage.2017.06.009.

[28] SCHMIDT P，GASER C，ARSIC M，et al. An

automated tool for detection of FLAIR-hyperintense

white-matter lesions in Multiple Sclerosis[J].

Neuroimage，2012，59（4）：3774-3783.