基于DCGAN的脑膜瘤与听神经瘤检测模型优化方法研究

doi:10.11938/cjmr20243127

摘要/Abstract

摘要：

由于人体桥小脑角区的脑膜瘤与听神经瘤在影像学的表现以及发病位置极其相似，临床诊断极易发生误诊．采用深度学习方法建立肿瘤自动检测模型，能有效降低人工诊断主观性，减少误诊漏诊率，提高工作效率．而数据集的多样性及图像质量的优越性很大程度上决定了检测模型的性能．针对医学图像数据集稀缺、类别数量不平衡及成像质量较差等问题，本文提出一种改进损失函数的深度卷积生成对抗网络（Deep Convolutional Generative Adversarial Networks，DCGAN）进行脑膜瘤与听神经瘤检测模型的数据增强，并与传统数据集增强方法进行了对比．结果显示通过改进的DCGAN优化数据集后，脑膜瘤与听神经瘤检测模型的精确率、特异性以及均值平均精度值（Mean Average Precision，mAP）分别较原数据集提高了0.014 6、0.022 4、0.030 0，上升至0.932 8、0.898 6、0.930 0．实验结果表明，通过DCGAN对数据集进行优化处理后，在脑肿瘤临床检测领域中，能较好地提高其模型的检测性能，更为可靠地辅助临床医学诊断．

关键词: 脑肿瘤, 检测模型, 数据集增强, DCGAN

Abstract:

Due to the extreme similarity in imaging manifestations and locations of onset between meningiomas and acoustic neuromas in the CPA (cerebellopontine angle) region of the human body, clinical diagnosis is prone to misdiagnosis. Establishing an automatic tumor detection model using deep learning methods can effectively reduce the subjectivity of manual diagnosis, decrease missed diagnosis rates, and improve work efficiency. The diversity of datasets and superiority of image quality largely determine the performance of the detection model. This paper proposes a DCGAN (deep convolutional generative adversarial networks) with improved loss function for data augmentation of meningioma and acoustic neuroma detection models to address the issues of scarce medical image datasets, imbalanced number of categories, and poor imaging quality. Compared with traditional dataset augmentation methods, the results show that after optimizing the dataset with DCGAN, the accuracy, specificity, and mAP (mean average precision) of the brain tumor detection model increase by 0.014 6, 0.022 4, and 0.030 0 respectively compared to the original dataset, reaching 0.932 8, 0.898 6, and 0.930 0. The study demonstrates that optimizing datasets with DCGAN can significantly improve the performance of the brain tumor detection model, providing a more reliable tool for clinical medical diagnosis.

Key words: brain tumors, detection model, dataset augmentation, DCGAN

中图分类号:

R739.41/O482.53

陈静聪, 冉凤伟, 章浩伟, 刘颖. 基于DCGAN的脑膜瘤与听神经瘤检测模型优化方法研究[J]. 波谱学杂志, 2025, 42(2): 117-129.

CHEN Jingcong, RAN Fengwei, ZHANG Haowei, LIU Ying. Optimization Methodology for Meningioma and Acoustic Neuroma Detection Model Based on DCGAN[J]. Chinese Journal of Magnetic Resonance, 2025, 42(2): 117-129.

图/表 13

图1

图2

表1

表2

图3

图4

图5

图6

图7

表3

图8

图9

图10

参考文献 28

[1]	RAN W B, LIANG Y C, QIN Q, et al. Medical image super-resolution reconstruction based on generative adversarial networks and attention mechanisms[J]. Intelligent Computer and Applications, 2023, 13 (1): 136-141.
	冉文兵, 梁永超, 覃芹, 等. 基于生成对抗网络和注意力机制的医学图像超分辨率重建[J]. 智能计算机与应用, 2023, 13(1): 136-141.
[2]	TAI Z Y, LI D D, LIU M. Medical image generation based on self attention mechanism and generative adversarial network[J]. Journal of Changchun University of Technology, 2024, 45 (3): 208-215.
	邰志艳, 李黛黛, 刘铭. 基于自注意力机制生成对抗网络的医学图像生成[J]. 长春工业大学学报, 2024, 45(3): 208-215.
[3]	GAN Y, YE M, ZENG F Y. A review of generative adversarial networks and their applications[J]. Journal of Chinese Computer Systems, 2020, 41(6): 1133-1139.
	淦艳, 叶茂, 曾凡玉. 生成对抗网络及其应用研究综述[J]. 小型微型计算机系统, 2020, 41(6): 1133-1139.
[4]	WU T Y, XU Y C, CHAO P F. Research on data enhancement based on generative adversarial networks[J]. Optics and Optoelectronic Technology, 2020, 18 (4): 47-52.
	吴天雨, 许英朝, 晁鹏飞. 基于生成对抗网络的数据增强研究[J]. 光学与光电技术, 2020, 18(4): 47-52.
[5]	WANG D X, QIN E Q, YUAN H C. Classification method of aquatic animals based on DCGAN data augmentation[J]. Fishery Modernization, 2019, 46(6): 68-75.
	王德兴, 秦恩倩, 袁红春. 基于DCGAN数据增强的水产动物分类方法[J]. 渔业现代化, 2019, 46(6): 68-75.
[6]	殷存军. 基于深度卷积网络和AdaBoost算法的无人机图像中病害松树识别[D]. 合肥: 安徽大学, 2020.
[7]	ZHAO N, SONG Y, YANG A, et al. Accurate classification of tunnel lining cracks using lightweight ShuffleNetV2-1.0-SE model with DCGAN-based data augmentation and transfer learning[J]. Appl Sci, 2024, 14(10): 4142.
[8]	ZHAO G, CAI Z, WANG X, et al. GAN Data augmentation methods in rock classification[J]. Appl Sci, 2023, 13(9): 5316.
[9]	DU X, DING X, XI M, et al. A data augmentation method for motor imagery EEG signals based on DCGAN-GP network[J]. Brain Sci, 2024, 14(4): 375.
[10]	HUANG H. Data augmentation by using gans and image transformation in facial emotion classification[C]// Journal of Physics: Conference Series. IOP Publishing, 2023, 2580(1): 012003.
[11]	MAYANK B, TRIPTI M. Comparison of affine and DCGAN-based data augmentation techniques for chest X-ray classification[J]. Procedia Comput Sci, 2023, 218: 283-290.
[12]	董家乐. 基于GAN的医学图像生成方法研究[D]. 郑州: 郑州大学, 2021.
[13]	于明浩. 基于深度卷积对抗生成网络的肺结节分类和分割方法[D]. 北京: 北京化工大学, 2021.
[14]	牟峙桦. 基于生成对抗网络的眼底图像生成方法研究[D]. 武汉: 华中科技大学, 2022.
[15]	SASWATI S, SUSHRUTA M, BAIDYANATH P, et al. An augmented modulated deep learning based intelligent predictive model for brain tumor detection using GAN ensemble[J]. Sensors-Basel, 2023, 23(15): 6930.
[16]	TIWARI A, HANNAN A S, PINNAMANENI R, et al. Optimized ensemble of hybrid RNN-GAN models for accurate and automated lung tumour detection from CT images[J]. Int J Adv Comput Sci Appl, 2023, 14(7): 621-631.
[17]	赵建峰. 基于生成对抗网络的肝血管瘤和肝细胞癌的增强及检测方法的研究[D]. 济南: 山东师范大学, 2020.
[18]	洪怡. 基于深度学习的脑肿瘤图像检测方法研究[D]. 长春: 长春工业大学, 2024.
[19]	HU X Y, LIU Y, CHEN S, et al. Identification of acoustic neuroma and meningioma in the cerebellopontine angle region using Mask RCNN fusion attention mechanism[J]. Chinese J Magn Reson, 2023, 40(3): 293-306.
	胡小洋, 刘颖, 陈淑, 等. 融合注意力机制Mask RCNN的桥小脑角区听神经瘤和脑膜瘤的识别研究[J]. 波谱学杂志, 2023, 40(3): 293-306. doi: 10.11938/cjmr20223045
[20]	LOU Y Z, LIU Y, JIANG H, et al. Research on classification algorithm for meningiomas and acoustic neuromas in the cerebellopontine angle region based on MRI and deep learning[J]. Chinese J Magn Reson, 2020, 37(3): 300-310.
	娄云重, 刘颖, 江华, 等. 基于MRI和深度学习的桥小脑角区脑膜瘤与听神经瘤分类算法研究[J]. 波谱学杂志, 2020, 37(3): 300-310. doi: 10.11938/cjmr20192753
[21]	LIU Y, CHEN J C, HU X Y, et al. Classification and localization of meningiomas and acoustic neuromas in the cerebellopontine angle region based on Mask RCNN[J]. Chinese J Magn Reson, 2021, 38(1): 58-68.
	刘颖, 陈静聪, 胡小洋, 等. 基于Mask RCNN的桥小脑角区脑膜瘤与听神经瘤分类定位研究[J]. 波谱学杂志, 2021, 38(1): 58-68. doi: 10.11938/cjmr20202825
[22]	WANG Z R, YANG J J, JIANG H N, et al. CNN training with twenty samples for crack detection via data augmentation[J]. Sensors, 2020, 20(17): 4849.
[23]	龙程. 基于对抗网络的图像数据集扩充研究与实现[D]. 西安: 西安理工大学, 2020.
[24]	KINGMA D P. Adam: A method for stochastic optimization[J]. arxiv preprint arxiv:1412.6980, 2014.
[25]	WANG L, JI X H, YANG M, etc. Mineral image recognition based on data augmentation and ensemble learning[J]. Geoscience Frontiers, 2024, 31(4): 87-94.
	王琳, 季晓慧, 杨眉, 等. 基于数据增强和集成学习的矿物图像识别[J]. 地学前缘, 2024, 31(4): 87-94. doi: 10.13745/j.esf.sf.2024.5.6
[26]	WUBINEH Z B, RUSIECKI A, HALAWA K. Classification of cervical cells from the pap smear image using the RESDCGAN data augmentation and ResNet50V2 with self-attention architecture[J]. Neural Comput Appl, 2024, 36: 1-15.
[27]	ONAKPOJERUO P E, MUSTAPHA T M, OZSAHIN U D, et al. A comparative analysis of the novel conditional deep convolutional neural network model, using conditional deep convolutional generative adversarial network-generated synthetic and augmented brain tumor datasets for image classification[J]. Brain Sci, 2024, 14(6): 559.
[28]	SARATH S, NAIR J J. Detection and classification of respiratory syndromes in original and modified DCGAN augmented neonatal infrared datasets[J]. Procedia Comput Sci, 2024, 233: 422-431.

网络层	卷积核	填充（padding）	步长（stride）	激活函数	BN算法
输入层	无	/	/	ReLU	采用
Conv1	5×5	1	2	ReLU	采用
Conv2	5×5	1	2	ReLU	采用
Conv3	5×5	1	2	ReLU	采用
Conv4	5×5	1	2	ReLU	采用
输出层	无	/	/	Tanh	不采用

网络层	卷积核	填充（padding）	步长（stride）	激活函数	BN算法
输入层	无	/	/	Leaky ReLu	不采用
Conv1	5×5	1	2	Leaky ReLu	采用
Conv2	5×5	1	2	Leaky ReLu	采用
Conv3	5×5	1	2	Leaky ReLu	采用
Conv4	5×5	1	2	Leaky ReLu	采用
输出层	无	/	/	Leaky ReLu	采用

Mask RCNN检测模型的主干网络	数据集增强方式	精确率	特异性	召回率	mAP
FPN+VGG16	原数据集	0.8537	0.8343	0.8832	0.7836
	传统数据集增强	0.8725	0.8563	0.8751	0.7982
	改进损失函数前的DCGAN数据集增强	0.8882	0.8623	0.8821	0.8167
	改进损失函数后的DCGAN数据集增强	0.9033	0.8715	0.8802	0.8255
FPN+VGG19	原数据集	0.9022	0.8691	0.8689	0.8312
	传统数据集增强	0.9095	0.8702	0.8658	0.8398
	改进损失函数前的DCGAN数据集增强	0.9105	0.8789	0.8593	0.8615
	改进损失函数后的DCGAN数据集增强	0.9156	0.8823	0.8502	0.8906
FPN+ResNet50	原数据集	0.9012	0.8589	0.8615	0.8000
	传统数据集增强	0.9078	0.8687	0.8589	0.8331
	改进损失函数前的DCGAN数据集增强	0.9089	0.8705	0.8575	0.8532
	改进损失函数后的DCGAN数据集增强	0.9123	0.8793	0.8536	0.8882
FPN+ResNet101	原数据集	0.9182	0.8762	0.8569	0.9000
	传统数据集增强	0.9235	0.8806	0.8523	0.9200
	改进损失函数前的DCGAN数据集增强	0.9253	0.8863	0.8511	0.9225
	改进损失函数后的DCGAN数据集增强	0.9328	0.8986	0.8563	0.9300