摘要
【目的】评价一种基于判别式模型学习的交互式分割软件在常规灰阶超声图像中分割肝脏局灶性病变(FLL)的价值。【方法】软件分别从用户标注出前景以及背景区域采集样本点,提取样本点特征,对病灶区和病灶边缘分别建立病灶区域判别模型和病灶边缘判别模型。最后,将上述两种模型合并为组合模型,得到最优的分割结果。建立FLL理想模型并收集45例FLL患者的常规灰阶超声图像共60幅检验软件的分割能力,将软件分割结果与高年资医师手工分割结果进行比较。【结果】以高年资医师手工分割结果作为金标准,软件分割所得区域的真阳性率(TP)为91.49%,假阳性率(FP)为14.22%,面积交叠率(AOM)为81.05%;分割所得结果的平均相对误差为0.04%,与手工分割间的差异无统计学意义(P>0.05)。【结论】基于判别式模型学习的交互式分割软件可以准确的在常规灰阶超声图像中分割出FLL。
【Objective】 To evaluate the performance of interactive segmentation software in focal liver lesions（FLL）on baseline gray scale ultrasound Images. 【Methods】 The software extracted positive and negative samples according to the foreground and background scribbles respectively and quantized candidate features. Then trained discriminative region model and discriminative boundary model respectively. Finally, combined the two histograms mentioned above in the form of a histogram to obtain the ideal result. Ideal FLL models were built and sixty ultrasound images of FLL from 45 patients were selected to test the software and compare the result with the manual segmentation by experienced radiologist. 【Results】 Taking the manual segmentation as the gold standard,the true positive ratio of segmentation software was 91.49%, the false positive ratio was 14.22%, the area overlap measure was81.05%.The average relative errors was 0.04%, and there was no statistical difference in performance between the gold standard and the software segmentation（P〈0.05）. 【Conclusion】 The interactive segmentation software based on discriminant model learning cansuccessfully segment FLL on baseline gray scale ultrasound images.
出处
《中山大学学报（医学科学版）》
CAS
CSCD
北大核心
2016年第1期131-136,共6页
Journal of Sun Yat-Sen University:Medical Sciences
基金
广东省教育部产学研结合项目(2012B091000101)
广东省产业技术研究与开发专项资金项目(2013B060500044)
关键词
超声
计算机辅助系统
交互式分割
肝脏局灶性病灶
ultrasound
computer assisted diagnosis
interactive segmentation
focal liver lesions
