基于迁移学习策略的肝纤维化分期诊断方法

doi:10.16180/j.cnki.issn1007-7820.2021.06.002

摘要/Abstract

摘要：

针对肝纤维化四分期准确率较低,S₂与S₃期分期难的问题,文中提出了一种基于迁移学习策略的肝纤维化诊断方法。该方法基于预训练好的深度残差网络模型,随机初始化各层权值参数,加入采用旋转、裁剪patch小块等方法扩充的数据集微调各类参数。经过softmax分类器结合patch小块投票原则,最终得到肝纤维化S₀~S₁、S₂、S₃、S₄期的分期准确率分别为93.75%、90.63%、87.50%、86.96%。该结果表明,文中方法在基于高频超声图像的肝纤维化定量诊断任务中达到了较好的效果。通过比较分析可知,文中所述方法优于其他已有方法,为临床计算机辅助诊断肝纤维化疾病提供了更加有效的解决方案。

关键词: 肝纤维化, 迁移学习, 深度残差网络, 微调, 投票原则, softmax分类器, 高频超声, 计算机辅助诊断

Abstract:

In view of the problem that the accuracy of classification of liver fibrosis into four stages is low, and the difficulties in distinguishing S₂ and S₃ stages, a method of liver fibrosis’s diagnosis based on transfer learning strategy is proposed in this study. This method is based on a pre-trained deep residual network model, randomly initializes the weight parameters of each layer, and adds a dataset expanded by the method of rotation and cutting patch blocks to fine-tune various parameters. After the softmax classifier combined with the patch block voting principle, the accuracy rates of liver fibrosis S₀~S₁、S₂、S₃ and S₄ stages are 93.75%、90.63%、87.50% and 86.96%, respectively. The proposed method achieves good results in the quantitative diagnosis of liver fibrosis based on high-frequency ultrasound images. The comparative analysis reveales that this method is superior to other existing methods, and provides a more effective solution for clinical computer-aided diagnosis of liver fibrosis diseases.

Key words: liver fibrosis, transfer learning, deep residual network, fine-tuning, voting principle, softmax classifier, high-frequency ultrasound, computer-aided diagnosis

中图分类号:

TP181

翟岳仙,刘翔,宋家琳. 基于迁移学习策略的肝纤维化分期诊断方法[J]. 电子科技, 2021, 34(6): 11-16.

ZHAI Yuexian,LIU Xiang,SONG Jialin. Staging Diagnosis of Liver Fibrosis Based on Transfer Learning Strategy[J]. Electronic Science and Technology, 2021, 34(6): 11-16.

图/表 10

图1

图2

图3

表1

ResNet-34网络结构"

层次名称	输出大小	34-layer
Conv1	32×32	7×7, 64, stride 2
Conv2_x	16×16	3×3 max pool, stride 2 $3 × 3,64 3 × 3,64$
Conv3_x	8×8	$3 × 3,128 3 × 3,128$
Conv4_x	4×4	$3 × 3,256 3 × 3,256$
Conv5_x	2×2	$3 × 3,512 3 × 3,512$
-	1×1	global poling, 4-d fc softmax

表1

图4

图5

图6

表2

表3

表4

其他网络模型分类性能对比试验"

方法名称	特点	肝纤维化阶段
方法名称	特点	S₀~ $S 1 /$ %	S₂/%	S₃/%	S₄/%
纹理分析法	GLCM参数+ SVM	62.50	68.75	66.67	73.91
文献[16]	肝包膜特征+ KNN/TSVM	68.75	62.50	54.17	65.22
ResNet-18	残差网络结构	87.50	84.34	83.33	86.96
本文方法	残差网络结构	93.75	90.63	87.50	86.96

表4

参考文献 16

[1]	陶晨阳, 郭津生, 史颖弘. 细胞外囊泡在肝纤维化疾病中的研究进展[J]. 中华肝脏病杂志, 2019,27(6):407-410.
	Tao Chenyang, Guo Jinsheng, Shi Yinghong. Research progress of extracellular vesicles in liver fibrosis[J]. Chinese Journal of Hepatology, 2019,27(6):407-410.
[2]	Huet N, Denis I, Martino A, et al. Ultrasonographic assessment of liver fibrosis with computer-assisted analysis of liver surface irregularities[J]. Diagn Interv Imaging, 2015,96(9):941-946. doi: 10.1016/j.diii.2015.02.010
[3]	伏广霞, 刘翔, 宋家琳, 等. 一种基于高频超声图像的肝包膜提取算法研究[J]. 电子科技, 2019,32(11):33-37.
	Fu Guangxia, Liu Xiang, Song Jialin, et al. Research on liver capsule extraction algorithm based on high-frequency ultrasound images[J]. Electronic Science and Technology, 2019,32(11):33-37.
[4]	陈明丽, 沈峰. 声像图纹理分析和人工神经网络在肝纤维化分期诊断中的初步应用[J]. 肝脏, 2013,18(8):541-543.
	Chen Mingli, Shen Feng. Preliminary application of sonographic texture analysis and artificial neural network in staging diagnosis of liver fibrosis[J]. Liver, 2013,18(8):541-543.
[5]	李宏彬, 赫光中, 程莉莉, 等. 基于CART决策树的B超影像肝纤维化纹理识别研究[J]. 中国数字医学, 2017,12(11):49-52.
	Li Hongbin, He Guangzhong, Cheng Lili, et al. Texture recognition of liver fibrosis in B-mode ultrasound imaging based on CART decision tree[J]. China Digital Medicine, 2017,12(11):49-52.
[6]	Lee W L. An ensemble-based data fusion approach for characterizing ultrasonic liver tissue[J]. Applied Soft Computing Journal, 2013,13(8):3683-3692. doi: 10.1016/j.asoc.2013.03.009
[7]	林健玲. 高频超声观察慢性肝炎肝实质形态对诊断肝纤维化的临床价值[J]. 临床肝胆病杂志, 2013,29(5):359-362.
	Lin Jianling. Clinical value of high-frequency ultrasound observation of liver parenchyma of chronic hepatitis in diagnosis of liver fibrosis[J]. Journal of Clinical Hepatobiliary Disease, 2013,29(5):359-362.
[8]	林萍, 王朝丽. 高频超声检测肝脏对肝纤维化分期的临床诊断价值[J]. 西北国防医学杂志, 2013,34(1):78-79.
	Lin Ping, Wang Chaoli. Clinical diagnosis of liver fibrosis staging by high-frequency ultrasound[J]. Journal of Northwest National Defence Medical Sciences, 2013,34(1):78-79.
[9]	翟岳仙, 刘翔, 宋家琳, 等. 基于超声的计算机辅助诊断肝纤维化研究进展[J]. 中国医学影像技术, 2019,35(7):1107-1110.
	Zhai Yuexian, Liu Xiang, Song Jialin, et al. Progress of computer-aided diagnosis of liver fibrosis based on ultrasonography[J]. Chinese Journal of Medicial Imaging Technology, 2019,35(7):1107-1110.
[10]	Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[C]. Qatar:International Conference on Neural Information Processing Systems, 2012.
[11]	Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]. Boston:IEEE Conference on Computer Vision and Pattern Recognition, 2014.
[12]	Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[C]. Banff: International Conference on Learning Representations, 2014.
[13]	He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]. Las Vegas:IEEE Conference on Computer Vision and Pattern Recognition, 2016.
[14]	Buslaev A, Seferbekov S, Iglovikov V, et al. Fully convolutional network for automatic road extraction from satellite imagery[C]. Salt Lake City:IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2018.
[15]	中华医学会传染病与寄生虫病学分会, 肝病学分会. 病毒性肝炎防治方案:中华医学会传染病与寄生虫病学分会、肝病学分会联合修[J]. 肝脏, 2000,5(4):257-263.
	Chinese Medical Association Infectious Diseases and Parasitology Branch, Hepatology Branch. Viral hepatitis prevention and treatment program: joint revision by the chinese medical association infectious and parasitology branch and hepatology branch[J]. Liver, 2000,5(4):257-263.
[16]	宋家琳, 刘翔, 章建全, 等. 肝包膜超声图像几何特征定量评估肝纤维化的实验研究[J]. 中华超声影像学杂志, 2018,27(12):86-90.
	Song Jialin, Liu Xiang, Zhang Jianquan, et al. Experimental study on quantitative evaluation of liver fibrosis with geometric features of liver capsule ultrasound images[J]. Chinese Journal of Ultrasound Imaging, 2018,27(12):86-90.