基于对抗生成网络的医学图像分析(论文13000字)
摘要:癌症目前是人类健康最大的威胁。自从数字病理和深度学习发展起来之后，研究人员逐渐的专注于开发智能计算机辅助诊断系统。目前常见的深度学习网络都是基于监督学习的训练，因此需要有大量的有标签数据。而在医学图像领域中，难以获得大量带标注的训练样本驱动深度网络，因此容易造成过拟合的问题。对抗生成网络(GAN)可以从少量的样本中学习到样本的分布特征，从而可以生成大量的同分布样本数据。所以为了解决过拟合这一矛盾性问题，本文提出了一种基于深度对抗网络的医学图像分析方法获取训练样本。首先根据少量的具有医学专家标记的样本训练对抗网络，然后对抗网络生成大量的伪造样本，最后基于Alexnet和lenet5做分类训练测试。本文提出的方法在三种医学数据及上测试：1)DCE-MRI核磁共振乳腺图像，原样本分类准确率62%，用对抗网络生成的样本作为训练集，用原始样本作为测试集得到的准确率达到了87%，用原始样本和对抗样本联合起来作为训练集，用原始图像和对抗生成的部分样本作为测试集，最后的准确率高达93.5%。2)口腔癌2类数据，原始样本训练准确率为90%，用GAN网络对抗生成正样本后训练准确率达到了98%。3)宫颈癌三类数据，原始数据训练测试得到的准确率为62.5%，经过GAN网络对三类数据集进行扩充后，训练测试的准确率达到了75%。
关键词：深度学习，对抗生成网络，医学图像分析
Medical Image Analysis Based on Generative Adversarial Networks
Abstract: Cancer is now the biggest threat to human health. Since the development of digital pathology and deep learning, researchers have gradually focused on the development of intelligent computer-aided diagnostic systems. At present, the common deep learning network is based on the supervision of the training, so the need for a large number of tagged data. In the field of medical images, it is difficult to obtain a large number of labeled training samples to drive deep network, so it is likely to cause over-fitting problems.The Generative Adversarial Networks(GAN) can learn the distribution characteristics of samples from a small number of samples, so that a large number of distributed data can be generated.So in order to solve this contradictory problem, this paper proposes a medical image analysis method based on GAN network to obtain training samples. Firstly according to a small number of medical experts with the sample training GAN, and then trained GAN generate a large number of forged samples, and finally based on Alexnet and lenet5 classification training test. The method proposed in this paper is based on three medical data and tests: 1) DCE-MRI nuclear mammography images, the original sample classification accuracy rate of 62%, with anti-network generated samples as a training set, with the original sample as a test set to get accurate The rate of 87%, with the original sample and confrontation samples together as a training set, with the original image and confrontation generated part of the sample as a test set, the final accuracy rate as high as 93.5%. 2) oral cancer 2 data, the original sample training accuracy rate of 90%, with GAN network against positive samples after the training accuracy rate reached 98%. 3) cervical cancer three types of data, the original data training test to obtain the accuracy rate of 62.5%, through the GAN network to expand the three types of data sets, the training test accuracy rate of 75%.
Key words: deep learning,Generative Adversarial Networks, medical image analysis
目录
摘要 1
1.引言 3
2.方法 6
2.1 生成对抗网络 6
2.2基于对抗生成网络的肿瘤分割模型 8
2.2.1 数据预处理 9
2.2.2 本文所采用的GAN（对抗生成网络）介绍 10
2.2.3 基于caffe的深度网络分类器简介 12
3. 数据集介绍 14
4. 实验 15
4．1实验步骤 15
4.1.1 三阴性核磁共振乳腺癌数据集 15
4.1.2 口腔癌数据集 16
4.1.3 宫颈癌数据集 16
4.2 实验结果 18
5.结论 20
参考文献 21
致谢 23
|
