摘要
提出一种局部三值模式方差(LTPV)描述符的特征提取方法,以提高现有胰腺内镜超声图像计算机辅助诊断系统在慢性胰腺炎与自身免疫性胰腺炎病例的鉴别诊断中的准确率。针对现有纹理特征区分度不足,对噪声鲁棒性较差的问题,该方法在局部二值模式描述符的基础上引入阈值参数,并在直方图建立过程中加入邻域对比度权重,最终得到10维特征向量。实验数据采用长海医院消化内科于2005年到2011年期间采集的内镜超声图像。结果表明,LTPV特征在SVM分类器下的各项分类性能都优于基于传统纹理特征和LBP特征的各类变种的分类结果。加入现有计算机辅助系统中后,分类准确性可达90.18%。可见,此特征对胰腺内镜超声图像具有较高的分类准确性和较好的鲁棒性,有望为自身免疫性胰腺炎与慢性胰腺炎的临床鉴别诊断提供有价值的参考。
A local ternary pattern variance(LTPV) based feature extraction method is proposed to improve the classification accuracy between chronic pancreatitis and autoimmune pancreatitis from endoscopic ultrasound images with the existing computer aided diagnosis(CAD) system. According to the problem of existing texture features, such as not enough discriminability and poor robustness to noise, the LTPV introduces the threshold parameter based on the local binary pattern and adds neighbor contrast weight into the histogram establishing procedure to get a 10 dimension feature vector. The proposed feature is evaluated on pancreas endoscopic ultrasound images collected by gastroenterology department of Changhaihospital from 2005 to 2011. The result indicates that the LTPV shows better classification capacity than other traditional features and varieties of the local binary pattern feature under SVM classifier. The accuracy can reach 90.18% after adding the proposed feature into the existing CAD system. The results suggest thatthe LTPV gains relatively higherclassification accuracy and better robustness on pancreas endoscopic ultrasound images.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2014年第S2期182-185,共4页
Chinese Journal of Scientific Instrument
基金
高等学校博士学科点专项科研基金(20110071110017)资助