Comparison Study of Radiomics and Deep Learning-Based Methods for Thyroid Nodules Classification Using Ultrasound Images
Comparison Study of Radiomics and Deep Learning-Based Methods for Thyroid Nodules Classification Using Ultrasound Images
复制标题
DOI：
10.1109/access.2020.2980290
复制
发表时间：
2020-01-01
期刊：
影响因子：
3.9
通讯作者：
Yang, Guang
中科院分区：
文献类型：
--
作者：
Wang, Yongfeng;Yue, Wenwen;Yang, Guang
甲状腺结节发病率高，恶性比例小。在物联网的帮助下，人们开发了许多非侵入性的方法来提高恶性结节的检出率。这些方法大致可以分为两类：基于放射组学的方法和基于深度学习的方法。总的来说，基于卷积神经网络的深度学习方法在许多医学图像分析和分类应用中取得了很好的表现；然而，目前还没有对基于放射组学和基于深度学习的方法进行比较。因此，在本文中，我们的目的是比较放射组学和基于深度学习的方法从超声图像中分类甲状腺结节的性能。一方面，我们开发了一种基于放射组学的方法，该方法包括从预处理图像中提取高通量的302维统计特征。然后分别利用互信息和线性判别分析进行降维，实现最终分类。另一方面，开发了一种基于深度学习的方法，并通过对VGG16模型进行预训练和微调进行了测试。回顾性收集1040例超声图像3120张，其中良性结节1841张，恶性结节1393张。数据集分为80 & x0025；培训和20 & x0025；测试数据。基于放射组学和深度学习的检测数据的最高准确率分别为66.81和x0025；和74.69 & x0025；,分别。对比结果表明，基于深度学习的方法比基于放射组学的方法具有更好的性能。
Thyroid nodules have a high prevalence and a small percentage is malignant. Many non-invasive methods have been developed with the help of the Internet of Things to improve the detection rate of malignant nodules. These methods can be roughly categorized into two classes: radiomics based and deep learning based approaches. In general, convolutional neural networks based deep learning methods have achieved promising performance in many medical image analysis and classification applications; however, no existing comparison has been done between radiomics based and deep learning based approaches. Therefore, in this paper, we aim to compare the performance of radiomics and deep learning based methods for the classification of thyroid nodules from ultrasound images. On one hand, we developed a radiomics based method, which consists of extracting high throughput 302-dimensional statistical features from pre-processed images. Then dimension reduction was performed using mutual information and linear discriminant analysis respectively to achieve the final classification. On the other hand, a deep learning based method was also developed and tested by pre-training a VGG16 model with fine-tuning. Ultrasound images including 3120 images (1841 benign nodules and 1393 malignant nodules) from 1040 cases were retrospectively collected. The dataset was divided into 80 & x0025; training and 20 & x0025; testing data. The highest accuracies yielded on the testing data for radiomics and deep learning based methods were 66.81 & x0025; and 74.69 & x0025;, respectively. A comparison result demonstrated that the deep learning based method can achieve a better performance than using radiomics.