一种基于完整结肠壁和纹理特征的小/平坦型息肉检测方法

Small/flat polyps detection based on texture features extracted from intact colon wall

目的:提出一种新的基于虚拟结肠镜的计算机辅助检测(computer-aided detection,CAD)方法,以提高对小/平坦型息肉的检测能力。方法:选择25例患者共89个息肉的CT数据作为训练集,6例患者共17个息肉的CT数据作为测试集。首先通过分割提取完整的结肠内、外壁;然后用直径9个体素的球体滑动窗在结肠内壁上逐点选取感兴趣体积(volume of interest,VOI)并计算267个纹理特征,用AdaBoost分类器进行分类预测得到初始疑似息肉(initial polyp candidates,IPCs);最后利用基于随机森林的特征选择策略进行特征寻优,并对疑似区域的每个体素进行分类得到最终的疑似息肉。结果:该方法对测试集检测敏感度为100%,假阳性率为38个/数据集。结论:该方法能够有效地检测小/平坦型息肉,提高传统CAD的性能。

Objective To propose a new computer-aided detection method based on virtual colonoscopy to enhance the examination of small/flat polyps.Methods Twenty-five patients with a total of 89 polyps had their CT data selected as the training set,and 6 ones with a total of 17 polyps had their CT data used as the test set.Firstly,the intact inner and outer walls of the colon were extracted by segmentation;then the volume of interest(VOI)was selected point by point on the inner wall of the colon using a spherical sliding window with a diameter of 9 elements and 267 textural features were calculated,and the initial polyp candidates(IPCs)were classified and predicted by AdaBoost classifier;finally,a random forest-based feature selection strategy was used for feature hunting and each voxel of the suspected volume was classified to obtain the final suspected polyp.Results The method had the sensitivity for the test set being 100%and the false positive rate being 38 per dataset.Conclusion The method proposed can effectively detect small/flat polyps and improve the performance of traditional CAD.