基于视觉信息与征象标签的肺结节CT图像检索

发布时间：2018-10-19 13:26

【摘要】：肺结节CT图像的相似性检索是计算机辅助诊断系统中最重要的部分,目前常用的检索方法通常匹配精度低,检索速度慢.针对上述问题,提出一种新的基于视觉信息与征象标签的双概率超图哈希算法,使用两层结构提高肺结节图像的检索精度:在第一层,将肺结节影像视觉信息和标签信息分别构建概率超图,最优划分概率超图得到哈希码;在第二层,使用结节图像的视觉特征、标签特征和第一层得到的哈希码来训练哈希函数.在检索时,对待检图像通过训练好的哈希函数进行0,1编码,与数据集中图像比较汉明距离,返回相似结节图像.对9种不同征象类型的3422张肺结节CT图像进行实验,并与不同哈希算法进行比较,结果表明,提出的方法在哈希码长为32位时可以达到最高精度90.18%,有效提高了检索精度,可以给医生提供客观的辅助诊断.
[Abstract]:The similarity retrieval of CT images of pulmonary nodules is the most important part in the computer-aided diagnosis system. At present, the commonly used retrieval methods usually have low matching accuracy and slow retrieval speed. In order to solve the above problems, a new double-probability hypergraph hashing algorithm based on visual information and sign label is proposed, which uses two-layer structure to improve the retrieval accuracy of pulmonary nodule image: in the first layer, In the second layer, the visual features of the nodule image, the label feature and the hash code of the first layer are used to train the hash function. In retrieval, the detected image is encoded by a trained hash function, and the hamming distance of the image is compared with that of the dataset image, and the similar nodule image is returned. 3422 pulmonary nodule CT images of 9 different types of signs were tested and compared with different hash algorithms. The results show that the proposed method can achieve the highest accuracy of 90.18 when the hash code length is 32 bits, and the retrieval accuracy is improved effectively. It can provide the doctor with objective auxiliary diagnosis.
【作者单位】：太原理工大学计算机科学与技术学院;山西省人民医院CT室;美国德克萨斯大学圣安东尼奥分校计算机科学系;
【基金】：国家自然科学基金(61373100) 山西省回国留学人员科研资助项目(2016-038) 虚拟现实技术与系统国家重点实验室开放基金(BUAA-VR-17KF-14,BUAA-VR-17KF-15)
【分类号】：TP391.41

【相似文献】