谭青,丁亚军,钱盛友,陈兴.基于超声识别HIFU致生物组织变性[J].中国医学影像技术,2020,36(6):913~917
基于超声识别HIFU致生物组织变性
In vitro recognition of HIFU-induced biological tissue degeneration based on ultrasound
投稿时间:2019-06-18  修订日期:2019-12-03
DOI:10.13929/j.issn.1003-3289.2020.06.028
中文关键词:  高强度超声聚焦疗法  组织  变性  识别  广义回归神经网络  体外研究
英文关键词:high intensity focused ultrasound therapy  tissue  degeneration  recognition  generalized regression neural network  in vitro
基金项目:国家自然科学基金(11774088)。
作者单位E-mail
谭青 湖南师范大学信息科学与工程学院, 湖南 长沙 410081  
丁亚军 湖南师范大学信息科学与工程学院, 湖南 长沙 410081 78865699@qq.com 
钱盛友 湖南师范大学物理与电子科学学院, 湖南 长沙 410081  
陈兴 湖南师范大学信息科学与工程学院, 湖南 长沙 410081  
摘要点击次数: 1387
全文下载次数: 515
中文摘要:
      目的 探讨以超声筛选最佳特征向量,结合广义回归神经网络(GRNN)识别强度聚焦超声(HIFU)致生物组织变性的方法。方法 采用HIFU以不同剂量对300个新鲜离体猪肉组织样本进行辐照,获得变性及未变性样本各150个。于辐照前后采集超声声像图,经减影处理获得超声减影图像;以灰度-梯度共生矩阵法及灰度差分统计法提取18个特征参数,经P值显著性检测法及欧氏距离法筛选获得最佳特向量。以300个样本中的198组为训练样本,102组为测试样本。识别训练样本后,以P值显著性检测法剔除的特征向量和欧氏距离最小的2个特征向量为最佳特征向量的对照组,将其分别输入GRNN,以识别组织变性;计算特征向量结合GRNN对测试样本的正确识别率和总识别率。结果 最佳特征向量为梯度分布不均匀性和灰度分布不均匀性,其结合GRNN的总识别率分别为90.20%、91.18%,以2个最佳特征组合并结合GRNN后总识别率为98.04%。P值显著性检测法剔除的特征向量为平均值、对比度,其结合GRNN的总识别率分别为48.04%、75.49%,2以2个最佳特征组合并结合GRNN的总识别率为79.41%。欧氏距离最小的特征向量为能量、小梯度优势,结合GRNN的总识别率分别为88.24%、89.22%,以2个最佳特征组合并结合GRNN的总识别率为89.22%。最佳特征向量组合结合GRNN可明显提高对变性组织的识别率。结论 基于超声减影图像,以灰度分布不均匀性、梯度分布不均匀性与GRNN结合,均可提高对HIFU辐照所致组织变性的识别率;以2个最佳特征组合结合GRNN识别效果更佳。
英文摘要:
      Objective To explore a method to improve the identification rate of tissue degeneration caused by high intensity focused ultrasound (HIFU) based on ultrasound combining with generalized regression neural network (GRNN). Methods Totally 300 fresh isolated pork tissue samples were selected and irradiated at different HIFU doses, then 150 denatured and 150 undenatured samples were obtained. Ultrasonic images of the samples were collected before and after irradiation, then ultrasonic subtraction images were obtained. A total of 18 characteristic parameters of ultrasonic subtractive images were extracted using gray-gradient co-occurrence matrix and gray difference statistical methods, and the best characteristic vectors were obtained with P-value significance detection method and Euclidean distance method. Among 300 samples, 198 were taken as training samples and 102 as test samples. After recognition of training samples, the feature vectors eliminated with P-value significance detection method and 2 feature vectors with the smallest Euclidean distance were taken as control group of the best feature vectors, and then were input into GRNN respectively for recognition of tissue denaturation. Correct recognition rate and total recognition rate of test samples were calculated using combining feature vectors with GRNN. Results The best feature vectors were non-uniformity of gray distribution and non-uniformity of gradient distribution, and the total recognition rate was 90.20% and 91.18% combining with GRNN, respectively, which increased to 98.04% when both 2 best characteristic parameters combined GRNN. The feature vectors eliminated using P-value significance detection method were average value and contrast, and the total recognition rate combining with GRNN was 48.04% and 75.49%, respectively, which became 79.41% when both 2 best characteristic parameters combined GRNN. The feature vectors with the smallest euclide distance were energy and small gradient, and the total recognition rate combining with GRNN was 88.24% and 89.22%, respectively, which remained 89.22% when both 2 of them combined with GRNN. The recognition rate of the optimal feature vectors combined with GRNN for tissue denaturation was significantly higher than that of control group. Conclusion Based on ultrasonic subtraction images, of pork tissue irradiated with HIFU, non-uniformity of gray distribution and non-uniformity of gradient distribution combined with GRNN can both improve the recognition rate of tissue denaturation, while the combination of them and GRNN is more effective in identifying tissue denaturation induced by HIFU.
查看全文  查看/发表评论  下载PDF阅读器