基于DenseNet结合迁移学习的胰腺囊性肿瘤分类方法

doi:10.11938/cjmr20223047

摘要/Abstract

摘要：

本文采用了DenseNet结合迁移学习的分类模型，对胰腺黏液性囊性肿瘤（MCN）和浆液性囊性肿瘤（SCN）进行分类. 首先对来自长海医院的65例MCN和107例SCN数据进行扩增和预处理，其次构建DenseNet结合迁移学习的分类模型并进行微调，实验过程采用五折交叉验证，对MCN和SCN进行识别分类，并将该模型与AlexNet、VGG16、ResNet50等其他深度学习模型进行对比. 结果显示本文的分类模型识别效果最好，在测试集上ROC曲线下面积（AUC值）达到0.989，准确率为0.943，召回率为0.949，精确率为0.938. 由此可见基于DenseNet结合迁移学习的分类模型对MCN和SCN具有较高的识别准确率，优于其他深度学习模型，并具有较强的学习能力，可以辅助医生在临床上的诊断，在一定程度上节省人力物力. 该模型对于胰腺囊性肿瘤识别分类的潜在价值和临床意义.

关键词: 深度学习, DenseNet, 迁移学习, 胰腺囊性肿瘤

Abstract:

This work applied the classification model of DenseNet combined with transfer learning to classify mucinous cystic tumor (MCN) from serous cystic tumor (SCN) of the pancreas. Firstly, the data of 65 MCNs and 107 SCNs from Changhai Hospital were augemented and preprocessed. Secondly, the classification model of DenseNet combined with transfer learning was constructed and fine-tuned, MCN and SCN were classified by 5-fold cross validation, and the proposed classification model was compared with AlexNet, VGG16, ResNet50 and other deep learning models. The classification model in this paper yielded the best recognition effect. The area under the ROC curve (AUC value), accuracy rate, recall rate and precision rate of the test set were 0.989, 0.943, 0.949 and 0.938 respectively. It proved that the classification model based on DenseNet combined with transfer learning has higher recognition accuracy for MCN and SCN and stronger learning ability than other deep learning models, which can help doctors in clinical diagnosis, and save manpower and material resources. It further confirmed the potential value and clinical significance of this model for the classification of pancreatic cystic tumors.

Key words: deep learning, DenseNet, transfer learning, cystic tumor of pancreas

中图分类号:

TP394.1

田慧, 武杰, 边云, 张志伟, 邵成伟. 基于DenseNet结合迁移学习的胰腺囊性肿瘤分类方法[J]. 波谱学杂志, 2023, 40(3): 270-279.

TIAN Hui, WU Jie, BIAN Yun, ZHANG Zhiwei, SHAO Chengwei. Classification of Pancreatic Cystic Tumors Based on DenseNet and Transfer Learning[J]. Chinese Journal of Magnetic Resonance, 2023, 40(3): 270-279.

图/表 8

表1

图1

表2

表3

表4

图2

图3

表5

参考文献 24

[1]	ARDESHNA D R, CAO T, RODGERS B, et al. Recent advances in the diagnostic evaluation of pancreatic cystic lesions[J]. World J Gastroenterol, 2022, 28(6): 624-634. doi: 10.3748/wjg.v28.i6.624
[2]	XU X B, CHEN H, SUN B. Progress in diagnosis and treatment of pancreatic cystic tumors[J]. Chinese Journal of Operative Procedures of General Surgery (Electronic Edition), 2020, 14(6): 643-646.
	徐西伯, 陈华, 孙备. 胰腺囊性肿瘤的诊治进展[J]. 中华普外科手术学杂志(电子版), 2020, 14(6): 643-646.
[3]	JANG D K, SONG B J, RYU J K, et al. Preoperative diagnosis of pancreatic cystic lesions: the accuracy of endoscopic ultrasound and cross-sectional imaging[J]. Pancreas, 2015, 44(8): 1329-1333. doi: 10.1097/MPA.0000000000000396
[4]	SUN Y, YANG S, QI E, et al. Comparative diagnostic evaluation with contrast-enhanced ultrasound, computed tomography and magnetic resonance imaging in patients with pancreatic cystic neoplasms[J]. Cancer Manag Res, 2020, 12: 2889-2898. doi: 10.2147/CMAR.S246564 pmid: 32425602
[5]	BOLLEN T L, WESSELS F J. Radiological workup of cystic neoplasms of the pancreas[J]. Visceral Med, 2018, 34(3): 182-190. doi: 10.1159/000489674
[6]	陈海燕. 影像学特征联合增强CT纹理特征鉴别胰腺浆液性囊腺瘤与黏液性囊腺瘤[D]. 浙江大学, 2020.
[7]	袁梦依. 基于影像组学特征融合的胰腺囊性肿瘤分类[D]. 浙江工业大学, 2020.
[8]	ZHANG Y F, XU S S, WU J, et al. Value of CT texture analysis in differentiating pancreatic serous cystadenoma from mucinous cystadenoma[J]. Journal of Southeast University (Medical Science Edition), 2022, 41(3): 308-316.
	张怡帆, 徐珊珊, 吴锦, 等. CT纹理分析在鉴别胰腺浆液性囊腺瘤与黏液性囊腺瘤中的价值[J]. 东南大学学报(医学版), 2022, 41(3): 308-316.
[9]	CHEN S, CHEN X, CHEN J Y, et al. Application of CT radiomics in differential diagnosis of pancreatic serous and mucinous cystic neoplasm[J]. Chinese Journal of CT and MRI, 2022, 20(10): 92-95.
	陈帅, 陈晓, 陈井亚, 等. 基于CT影像组学对胰腺浆液及黏液性囊性肿瘤鉴别诊断[J]. 中国CT和MRI杂志, 2022, 20(10): 92-95.
[10]	YANG Y F, QI Z X, NIE S D. Differentiation of benign and malignant breast lesions based on multimodal MRI and deep learning[J]. Chinese J Magn Reson, 2022, 39(4): 401-412.
	杨一风, 祁章璇, 聂生东. 基于多模态MRI与深度学习的乳腺病变良恶性鉴别[J]. 波谱学杂志, 2022, 39(4): 401-412.
[11]	WEI Z H, YAN S J, HAN B S, et al. Diagnosis of alzheimer's disease based on multi-output three-dimensional convolutional neural network[J]. Chinese J Magn Reson, 2021, 38(1): 92-100.
	魏志宏, 闫士举, 韩宝三, 等. 基于多输出的3D卷积神经网络诊断阿尔兹海默病[J]. 波谱学杂志, 2021, 38(1): 92-100.
[12]	NGUON L S, SEO K, LIM J H, et al. Deep learning-based differentiation between mucinous cystic neoplasm and serous cystic neoplasm in the pancreas using endoscopic ultrasonography[J]. Diagnostics, 2021, 11(6): 1052. doi: 10.3390/diagnostics11061052
[13]	VILAS-BOAS F, RIBEIRO T, AFONSO J, et al. Deep learning for automatic differentiation of mucinous versus non-mucinous pancreatic cystic lesions: a pilot study[J]. Diagnostics, 2022, 12(9): 2041. doi: 10.3390/diagnostics12092041
[14]	HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]// Proceedings of the IEEE conference on computer vision and pattern recognition, 2017: 4700-4708.
[15]	ZHOU Q, ZHOU Z, CHEN C, et al. Grading of hepatocellular carcinoma using 3D SE-DenseNet in dynamic enhanced MR images[J]. Comput Biol Med, 2019, 107: 47-57. doi: S0010-4825(19)30032-0 pmid: 30776671
[16]	WANG Z, LI X, YAO M, et al. A new detection model of microaneurysms based on improved FC-DenseNet[J]. Sci Rep, 2022, 12(1): 1-9. doi: 10.1038/s41598-021-99269-x
[17]	GUO W, XU Z, ZHANG H. Interstitial lung disease classification using improved DenseNet[J]. Multimed Tools Appl, 2019, 78(21): 30615-30626. doi: 10.1007/s11042-018-6535-y
[18]	YANG Y H, LIU M, WANG X M, et al. Breast cancer image recognition based on DenseNet and transfer learning[J]. Journal of Jilin University, 2022, 40(2): 213-218.
	杨雨航, 刘铭, 王新民, 等. 基于DenseNet和迁移学习的乳腺癌图像识别[J]. 吉林大学学报, 2022, 40(2): 213-218.
[19]	ZHOU Y, ZHANG X, WANG Y, et al. Transfer learning and its application research[J]. J Phys: Conf Ser, 2021, 1920(1): 012058. doi: 10.1088/1742-6596/1920/1/012058
[20]	LI Y, SONG P H. Review of transfer learning in medical image classification[J]. Journal of Image and Graphics, 2022, 27(3): 672-686.
	黎英, 宋佩华. 迁移学习在医学图像分类中的研究进展[J]. 中国图象图形学报, 2022, 27(3): 672-686.
[21]	KINGMA D P, BA J. Adam: A method for stochastic optimization[J]. arXiv preprint arXiv:1412.6980, 2014.
[22]	ZEILER M D. Adadelta: an adaptive learning rate method[J]. arXiv preprint arXiv:1212.5701, 2012.
[23]	DUCHI J, HAZAN E, SINGER Y. Adaptive subgradient methods for online learning and stochastic optimization[J]. J Mach Learn Res, 2011, 12(7): 2121-2159.
[24]	DEAN J, CORADO G, MONGA R, et al. Large scale distributed deep networks[J]. Adv Neural Inf Process Syst, 2012, 25: 1-11.

Layers	DenseNet161
Convolution	7×7 conv
Pooling	3×3 max pool
Dense Block 1	$\left[ \begin{matrix} 1\times 1\ \text{conv} \\ 3\times 3\ \text{conv} \\ \end{matrix} \right]\times 6$
Transition Layer 1	1×1 conv, 2×2 average pool
Dense Block 2	$\left[ \begin{matrix} 1\times 1\ \text{conv} \\ 3\times 3\ \text{conv} \\ \end{matrix} \right]\times 12$
Transition Layer 2	1×1 conv, 2×2 average pool
Dense Block 3	$\left[ \begin{matrix} 1\times 1\ \text{conv} \\ 3\times 3\ \text{conv} \\ \end{matrix} \right]\times 36$
Transition Layer 3	1×1 conv, 2×2 average pool
Dense Block 4	$\left[ \begin{matrix} 1\times 1\ \text{conv} \\ 3\times 3\ \text{conv} \\ \end{matrix} \right]\times 24$
Classification Layer	7×7 global average pool, fully-connected, softmax

模型	ACC	AUC
DenseNet161	0.767	0.837
DenseNet161+迁移学习	0.943	0.989

模型	ACC	Precision	Recall	Specificity	F1-Score	AUC	参数量
AlexNet	0.844	0.846	0.842	0.847	0.844	0.930	6100840
MobileNet_v3	0.889	0.894	0.884	0.895	0.889	0.960	5483032
Vgg16	0.898	0.907	0.886	0.909	0.896	0.966	138365992
ResNet50	0.912	0.923	0.898	0.926	0.910	0.971	25557032
本文	0.943	0.938	0.949	0.937	0.943	0.989	28681000

	方法	数据	ACC
文献[7]	影像组学特征融合	胰腺囊性肿瘤CT图像	0.862
文献[8]	CT纹理特征模型+临床影像学特征模型	SCN与MCN静脉期CT图像	0.938 (AUC)
文献[9]	影像组学参数+人工神经网络	SCN与MCN CT图像	0.895
文献[12]	ResNet50+迁移学习	SCN与MCN内窥镜超声图像	0.828
本文	DenseNet161+迁移学习	SCN与MCN T₂加权MR图像	0.943