3D-GA-Unet：基于3D-Ghost卷积神经网络的脑胶质瘤MRI图像分割算法

doi:10.11772/j.issn.1001-9081.2023050606

摘要/Abstract

摘要： 脑胶质瘤是由于大脑和脊髓胶质癌变产生的、最常见的原发性颅脑肿瘤，其中恶性脑胶质瘤占比大且死亡率高。利用MRI图像对脑胶质瘤定量分割和分级是目前诊治脑胶质瘤的主要方法。为了提升脑胶质瘤的分割精度与速率，提出了一种基于3D-Ghost卷积神经网络的脑胶质瘤MRI图像分割模型：3D-GA-Unet。该模型以3D-Unet为基础，设计了基于空间注意力机制的3D-ghost卷积神经网络模块，利用线性运算增加有用信息输出、减少传统卷积神经网络中的冗余特征；添加了基于空间的注意力机制模块，有利于获取更多于分割精度有利的图像信息；在公共脑胶质瘤数据集BraTS2018进行训练和验证。实验结果表明3D-GA-Unet脑胶质瘤分割结果的水肿区域、核心区域和增强区域的平均Dice Score分别达到0.86，0.85，0.80，平均敏感度分别达到0.86，0.94，0.83。本文设计网络模块能够对脑胶质瘤图像进行精确分割，进一步提升分割效率，对脑胶质瘤的临床诊断有积极的意义。

关键词: 脑胶质瘤, 医学图像分割, 神经网络, 注意力机制, 卷积神经网络, U-net

Abstract: Glioma is the most common primary cranial tumor, of which malignant glioma accounts for a large proportion and has a high mortality rate. The quantitative segmentation and grading of gliomas using multimodal magnetic resonance imaging (MRI) images is the main method for the diagnosis and treatment of gliomas. In order to improve the segmentation accuracy of glioma and reduce the computational overhead, a multimodal glioma MRI image segmentation model based on 3D-Ghost convolutional neural network is innovatively proposed: 3D-GA-Unet. Based on 3D-Unet, this model designs a 3D-ghost convolutional neural network module based on spatial attention mechanism to improve the efficiency of feature extraction; use linear operations to increase useful information output and reduce redundant features in traditional convolutional neural networks; this paper proposes an image description model that combines multi-level decoder and dynamic fusion mechanism, which is an extension of the standard encoder-decoder structure, which can solve the gradient disappearance phenomenon that is easy to occur by traditional cascading structures, and make model training more stable; an BECDice loss function combining Dice Similariy Coefficient (DSC) and binary Cross Entropy (loss function) is proposed to solve the problem of unclear glioma boundaries in brain MRI images; the addition of a space-based attention mechanism module is conducive to obtaining more image information that is beneficial to segmentation accuracy. Training and validation were performed in the public glioma datasets BraTS2018, and compared with three other image segmentation models. The experimental results show that the average Dice Score of the edema area, the core area and the enhanced region of the 3D-GA-Unet glioma segmentation results reached 0.86, 0.85, 0.80, and the average sensitivity reached 0.86, 0.94, 0.83, respectively. The thesis design network module can accurately segment glioma images under limited overhead, further improve the segmentation efficiency, and have positive significance for the clinical diagnosis of glioma.

Key words: glioma, medical image segmentation, neural network, attention mechanism, CNN, U-net

中图分类号:

TP391

许立君黎辉刘祖阳陈侃松马为駽. 3D-GA-Unet：基于3D-Ghost卷积神经网络的脑胶质瘤MRI图像分割算法[J]. 计算机应用, DOI: 10.11772/j.issn.1001-9081.2023050606.

图/表 14

参考文献 32

1	OSTROM Q T， PATIL N， CIOFFI G， et al. CBTRUS Statistical Report： primary brain and other central nervous system tumors diagnosed in the United States in 2013—2017［J］. Neuro-Oncology， 2020， 22（S1）： iv1-iv96. 10.1093/neuonc/noaa200
2	STUPP R， BRADA M， VAN DEN BENT M J， et al. High-grade malignant glioma： ESMO clinical practice guidelines for diagnosis， treatment and follow-up ［J］. Annals of Oncology， 2014， 25（3）： 93-101. 10.1093/annonc/mdu050
3	LAPOINTE S， PERRY A， BUTOWSKI N A. Primary brain tumours in adults［J］.Lancet， 2018， 392（10145）： 432-446. 10.1016/s0140-6736(18)30990-5
4	WESSELING P， CAPPER D. WHO 2016 classification of gliomas［J］. Neuropathology and Applied Neurobiology， 2018， 44（2）： 139-150. 10.1111/nan.12432
5	李洁，刘光耀，樊凤仙，等. 深度学习和影像组学在脑胶质瘤中的研究进展［J］. 磁共振成像， 2022， 13（4）：158-161. 10.12015/issn.1674-8034.2022.04.035
	LI J， LIU G Y， FAN F X， et al. Advances in deep learning and imagingomics in glioma［J］. Magnetic Resonance Imaging， 2022 13（4）： 158-161. 10.12015/issn.1674-8034.2022.04.035
6	姜迪，刘慧，李钰，等. 结合稠密特征映射的CT图像肿瘤分割模型［J］. 计算机辅助设计与图形学学报， 2021，33（8）： 1273-1286. 10.3724/sp.j.1089.2021.18656
	JIANG D， LIU H， LI Y， et al. CT image tumor segmentation model combined with dense feature mapping［J］. Journal of Computer-Aided Design & Computer Graphics， 2021， 33（8）： 1273-1286. 10.3724/sp.j.1089.2021.18656
7	CLAES A， IDEMA A J， WESSELING P. Diffuse glioma growth： a guerilla war［J］. Acta Neuropathologica， 2007， 114（5）： 443-458. 10.1007/s00401-007-0293-7
8	BELAROUSSI B， MILLES J， CARME S， et al. Intensity non-uniformity correction in MRI： existing methods and their validation［J］. Medical Image Analysis， 2006， 10（2）： 234-246. 10.1016/j.media.2005.09.004
9	FRID-ADAR M， DIAMANT I， KLANG E， et al. GAN-based synthetic medical image augmentation for increased cnn performance in liver lesion classification ［J］. Neurocomputing， 2018， 321： 321-331. 10.1016/j.neucom.2018.09.013
10	SHELHAMER E， LONG J， DARRELL T. Fully convolutional networks for semantic segmentation［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2017， 39（4）： 640-651. 10.1109/tpami.2016.2572683
11	NOH H， HONG S， HAN B. Learning deconvolution network for semantic segmentation ［C］// Proceedings of the 2015 IEEE International Conference on Computer Vision. Piscataway： IEEE， 2015：1520-1528. 10.1109/iccv.2015.178
12	CHEN C， LIU X， DING M， et al. 3D dilated multi-fiber network for real-time brain tumor segmentation in MRI［C］// Proceedings of the 2019 International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham： Springer， 2019： 184-192. 10.1007/978-3-030-32248-9_21
13	CHEN L-C， PAPANDREOU G， KOKKINOS I， et al. DeepLab： semantic image segmentation with deep convolutional nets， atrous convolution， and fully connected CRFs［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2018， 40（4）： 834-848. 10.1109/tpami.2017.2699184
14	BADRINARAYANAN V， KENDALL A， CIPOLLA R. SegNet： a deep convolutional encoder-decoder architecture for image segmentation［J］. IEEE Transactions on Pattern Analysis & Machine Intelligence， 2017， 39（12）： 2481-2495. 10.1109/tpami.2016.2644615
15	DONG H， YANG G， LIU F， et al. Automatic brain tumor detection and segmentation using U-Net based fully convolutional networks［C］// Proceedings of the 2017 Annual Conference on Medical Image Understanding and Analysis. Cham： Springer， 2017： 506-517. 10.1007/978-3-319-60964-5_44
16	王海鸥，刘慧，郭强，等. 面向医学图像分割的超像素U-Net网络设计［J］. 计算机辅助设计与图形学学报， 2019， 31（6）：1007-1017. 10.3724/sp.j.1089.2019.17389
	WANG H O， LIU H， GUO Q， et al. Design of superpiexl U-Net network for medical image segmentation［J］. Journal of Computer-Aided Design & Computer Graphics， 2019，31（6）：1007-1017. 10.3724/sp.j.1089.2019.17389
17	ZHOU S， NIE D， ADELI E， et al. High-resolution encoder-decoder networks for low-contrast medical image segmentation［J］. IEEE Transactions on Image Processing， 2020， 29： 461-475. 10.1109/tip.2019.2919937
18	VALANARASU J M J， SINDAGI V A， HACIHALILOGLU I， et al. KiU-Net： overcomplete convolutional architectures for biomedical image and volumetric segmentation［J］. IEEE Transactions on Medical Imaging， 2022， 41（4）： 965-976. 10.1109/tmi.2021.3130469
19	QIN X， ZHANG Z， HUANG C， et al. U²-Net： going deeper with nested U-structure for salient object detection ［J］. Pattern Recognition， 2020， 106：107404. 10.1016/j.patcog.2020.107404
20	AHMED S F， RAHMAN F S， TABASSUM T， et al. 3D U-Net： fully convolutional neural network for automatic brain tumor segmentation［C］// Proceedings of the 2019 22nd International Conference on Computer and Information Technology. Piscataway： IEEE， 2019： 1-6. 10.1109/iccit48885.2019.9038237
21	AboELENIN N M， PIAO S， NOOR A， et al. MIRAU-Net： an improved neural network based on U-Net for gliomas segmentation［J］. Signal Processing： Image Communication， 2022， 101： 116553. 10.1016/j.image.2021.116553
22	WEN X， ZHAO B， YUAN M， et al. Application of multi-scale fusion attention U-Net to segment the thyroid gland on localized computed tomography images for radiotherapy［J］. Frontiers in Oncology， 2022， 12： 844052. 10.3389/fonc.2022.844052
23	YANG S， ZHOU X， WANG J， et al. Unsupervised domain adaptation for cross-device OCT lesion detection via learning adaptive features［C］// Proceedings of the 2020 IEEE 17th International Symposium on Biomedical Imaging. Piscataway： IEEE， 2020： 1570-1573. 10.1109/isbi45749.2020.9098380
24	PEIRIS H， HAYAT M， CHEN Z L， et al. A robust volumetric transformer for accurate 3D tumor segmentation［C］// Proceedings of the 2022 International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham： Springer， 2022： 162-172. 10.1007/978-3-031-16443-9_16
25	LI G， ZHANG H， E L， et al. Recognition of honeycomb lung in CT images based on improved MobileNet model［J］. Medical Physics， 2021， 48（8）：4304-4315. 10.1002/mp.14873
26	ZHANG X， ZHOU X， LIN M， et al. ShuffleNet： an extremely efficient convolutional neural network for mobile devices［C］// Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington， DC： IEEE Computer Society， 2018：6848-6856. 10.1109/cvpr.2018.00716
27	HAN K， WANG Y， QI T， et al. GhostNet： more features from cheap operations［C］// Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington， DC： IEEE Computer Society， 2020：1577-1586. 10.1109/cvpr42600.2020.00165
28	HOU Q， ZHAN D， FENG J. Coordinate attention for efficient mobile network design［C］// Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington， DC： IEEE Computer Society， 2021： 13708-13717. 10.1109/cvpr46437.2021.01350
29	MENZE B H， JAKAB A， BAUER S， et al. The multimodal Brain Tumor Image Segmentation Benchmark （BRATS）［J］. IEEE Transactions on Medical Imaging， 2015， 34（10）： 1993-2024. 10.1109/tmi.2014.2377694
30	WANG G， LI W， OURSELIN S， et al. Automatic brain tumor segmentation based on cascaded convolutional neural networks with uncertainty estimation ［J］. Frontiers in Computational Neuroscience， 2019， 13： No.56. 10.3389/fncom.2019.00056
31	ZHOU Z， HE Z， JIA Y. AFPNet： a 3D fully convolutional neural network with atrous-convolution feature pyramid for brain tumor segmentation via MRI images ［J］. Neurocomputing， 2020，402： 235-244. 10.1016/j.neucom.2020.03.097
32	ZHANG D， HUANG G， ZHANG Q， et al. Exploring task structure for brain tumor segmentation from multi-modality MR images［J］. IEEE Transactions on Image Processing， 2020， 29： 9032-9043. 10.1109/tip.2020.3023609

区域名	包含真实区域	掩码值	颜色	说明
WT	周围水肿区域（ED）	2	绿色	边界包含脑胶质瘤的所有肿瘤结构
ET	增强肿瘤区域（ET）	4	黄色	通过造影剂显示的肿瘤区域
TC	非增强肿瘤核（NET）坏死核心（NCR）	1	红色	脑胶质瘤的核心部分，包含主要的肿瘤结构
背景	无	0	黑色	图像中脑胶质瘤区域外的其他区域

区域名	包含真实区域	掩码值	颜色	说明
WT	周围水肿区域（ED）	2	绿色	边界包含脑胶质瘤的所有肿瘤结构
ET	增强肿瘤区域（ET）	4	黄色	通过造影剂显示的肿瘤区域
TC	非增强肿瘤核（NET）坏死核心（NCR）	1	红色	脑胶质瘤的核心部分，包含主要的肿瘤结构
背景	无	0	黑色	图像中脑胶质瘤区域外的其他区域

模型	DSC			PPV			sensitivity			Hausdorff			参数量/10⁶	收敛时间/h
模型	WT	TC	ET	WT	TC	ET	WT	TC	ET	WT	TC	ET	参数量/10⁶	收敛时间/h
U-Net	0.838 7	0.821 5	0.768 8	0.856 1	0.853 3	0.774 5	0.857 0	0.901 4	0.829 5	2.662 3	1.751 1	2.846 6	16.323 1	5.08
GA-Unet	0.841 3	0.836 6	0.774 2	0.872 8	0.863 7	0.804 1	0.861 3	0.905 8	0.825 8	2.596 9	1.717 4	2.807 6	16.323 4	5.30
3D U-Net	0.856 2	0.851 7	0.802 3	0.885 1	0.867 7	0.815 4	0.866 9	0.914 3	0.838 6	2.373 9	1.615 3	2.704 1	16.622 9	4.94
3D-GA-Unet	0.863 2	0.847 3	0.803 6	0.867 3	0.826 9	0.983 1	0.867 6	0.949 2	0.831 5	2.079 3	1.601 2	2.573 8	16.333 0	4.50

模型	DSC			PPV			sensitivity			Hausdorff			参数量/10⁶	收敛时间/h
模型	WT	TC	ET	WT	TC	ET	WT	TC	ET	WT	TC	ET	参数量/10⁶	收敛时间/h
U-Net	0.838 7	0.821 5	0.768 8	0.856 1	0.853 3	0.774 5	0.857 0	0.901 4	0.829 5	2.662 3	1.751 1	2.846 6	16.323 1	5.08
GA-Unet	0.841 3	0.836 6	0.774 2	0.872 8	0.863 7	0.804 1	0.861 3	0.905 8	0.825 8	2.596 9	1.717 4	2.807 6	16.323 4	5.30
3D U-Net	0.856 2	0.851 7	0.802 3	0.885 1	0.867 7	0.815 4	0.866 9	0.914 3	0.838 6	2.373 9	1.615 3	2.704 1	16.622 9	4.94
3D-GA-Unet	0.863 2	0.847 3	0.803 6	0.867 3	0.826 9	0.983 1	0.867 6	0.949 2	0.831 5	2.079 3	1.601 2	2.573 8	16.333 0	4.50

模型	DSC			PPV			sensitivity			Hausdorff			参数量/10⁶	收敛迭代次数	收敛时间/h
模型	WT	TC	ET	WT	TC	ET	WT	TC	ET	WT	TC	ET	参数量/10⁶	收敛迭代次数	收敛时间/h
3D U-Net	0.856 2	0.851 7	0.802 3	0.885 1	0.867 7	0.815 4	0.866 9	0.914 3	0.838 6	2.373 9	1.615 3	2.704 1	16.322 9	84	4.94
3D U-Net+Ghost	0.857 3	0.840 3	0.793 2	0.863 4	0.868 0	0.834 1	0.862 0	0.915 8	0.843 2	2.341 2	1.605 1	2.664 2	16.322 7	82	4.92
3D U-Net+Attention	0.858 2	0.853 2	0.801 6	0.875 2	0.841 2	0.842 4	0.861 4	0.924 3	0.821 3	2.213 4	1.612 1	2.584 1	16.334 6	79	4.81
3D-GA-Unet	0.863 2	0.847 3	0.803 6	0.867 3	0.826 9	0.983 1	0.867 6	0.949 2	0.831 5	2.079 3	1.601 2	2.573 8	16.333 0	69	4.50