胎盘成熟度自动分级探索

发布时间：2018-06-18 06:47

本文选题：胎盘成熟度评估 + 特征融合　；参考：《深圳大学》2017年硕士论文

【摘要】：胎盘成熟度分级的准确性对小于胎龄儿、死胎及死产的临床诊断有重要的作用。但是由于成像过程复杂、妊娠期长、图像质量差异以及医生主观判断差异,导致胎盘成熟度分级成为耗时又冗长的工作。尽管有近年来医学成像手段和技术都有了很大的进步,胎盘成熟度准确分级仍然有很大的挑战性。为了解决这一问题,本文提出一种基于特征融合和判别式学习的胎盘成熟度自动分级算法。首先,从B型灰阶超声图像中提取灰度信息,同时从彩色多普勒能量(color Doppler energy imaging,CDE)超声图像中提取血管信息,通过高斯差分的方法提取关键点,并对关键点提取尺度不变特征变换(scale-invariant feature transform,SIFT)特征以及灰度特征,对两种特征进行连接融合,利用Fisher向量编码的方法,形成码书,经过归一化,最后用支持向量机(support vector machine,SVM)进行分类,最终得到了 92.7%的准确率。对比不同关键点检测方法、不同特征提取方法以及不同的特征编码方法所得的实验结果表明,本文所提出的方法在胎盘成熟度自动分级问题中能取得很好的结果,对临床的判断有一定的指导意义。深度学习的发展为我们进一步提升结果的准确率提供了可能。有限的特征描述不能完整表述图像信息,所以本文通过将现有主流卷积神经网络(convolutional neural network,CNN)方法用于解决胎盘成熟度自动分级问题,用卷积层提取特征,并以池化层加速计算,以端到端的方式得到分类结果,减少人为对特征的选择和干预,得到更好的结果。本文以先前所取得的数据做训练,再以此后采集的数据为测试数据,利用AlexNet、VGG-F、VGG-S、VGG-M以及VGG-VD-16和VGG-VD-19进行实验,得到了相较于传统机器学习方法更好的结果,为进一步设计适用于该问题的网络结构提供了基础。本文在以传统的机器学习方法解决胎盘成熟度自动分级问题中,首次引入特征融合的思想,并在特征编码中采用多层Fisher向量编码,增强局部特征,从而提升分类的准确率。后续研究中对流行的深度学习方法的尝试,并取得了不错的结果,也为胎盘成熟度自动分级算法的临床应用提供了新的思路。
[Abstract]:The accuracy of placental maturity classification plays an important role in the clinical diagnosis of small gestational age, stillbirth and stillbirth. However, due to the complexity of imaging process, the long gestation period, the difference of image quality and the difference of doctors' subjective judgment, placental maturity grading becomes a time-consuming and lengthy task. Although great progress has been made in medical imaging techniques and techniques in recent years, accurate classification of placental maturity remains a challenge. To solve this problem, an automatic placental maturity classification algorithm based on feature fusion and discriminant learning is proposed. Firstly, grayscale information is extracted from B-type gray scale ultrasound image, and vascular information is extracted from color Doppler energy image. The key points are extracted by Gao Si difference method. The scale-invariant feature transform sift feature and gray scale feature are extracted from the key points. The two features are joined and fused. The codebook is formed by using the Fisher vector coding method. After normalization, the support vector machine is used to classify the two features. Finally, the accuracy rate of 92.7% was obtained. Compared with different key point detection methods, different feature extraction methods and different feature coding methods, the experimental results show that the proposed method can achieve good results in the placental maturity automatic grading problem. It has certain guiding significance to clinical judgment. The development of deep learning makes it possible for us to further improve the accuracy of the results. The finite feature description can not represent the image information completely, so we use the existing convolutional neural network to solve the problem of automatic grading of placenta maturity, extract the feature with convolution layer, and calculate it with pool layer. The classification results are obtained by end-to-end approach, and better results are obtained by reducing the artificial selection and intervention of features. In this paper, we use the data obtained before for training, and then take the data collected since then as the test data. We use AlexNet VGG-FGG-FG VGG-SGG-M, VGG-VD-16 and VGG-VD-19 to carry out experiments. The results are better than those of traditional machine learning methods. It provides the foundation for further designing the network structure suitable for this problem. In this paper, the idea of feature fusion is introduced for the first time in the traditional machine learning method to solve the placental maturity automatic grading problem, and the multi-layer Fisher vector coding is used in the feature coding to enhance the local features, thus improving the accuracy of classification. In the following research, the popular deep learning method has been tried, and good results have been obtained. It also provides a new idea for clinical application of placental maturity automatic grading algorithm.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R714.5;R445.1

【相似文献】