目的:提出并设计使用PET-CT影像定位肺部病灶区域并辅助判断病灶点的自动化流程,并对整个算法流程进行评价和分析,以提高临床工作效率.方法:选取北京协和医院核医学科20例肺部肿瘤患者的PET-CT影像,使用图像尺度变换等图像处理方法,去除CT图像中的床位,用等高轮廓线在PET-CT影像中提取样本区域,并依据预先标记的病灶区域信息对样本区域进行类别划分,提取每个样本区域的图像特征.应用Adaboost算法进行训练,建立相应的分类模型,利用训练好的分类模型对测试集进行测试,对比弱分类器构成的集成分类模型的准确率,用检出率、误检率、感兴趣区域(ROC)曲线以及病例分类的正确率对分类结果进行评估.结果:对20例患者的PET-CT图像预处理后,共产生125088个样本,其中正样本22720个,负样本为102368个,用等高轮廓线进行区域划分,使用Adaboost.M2算法融合多种特征训练出来的强分类器的样本分类正确率为97%左右,20例肺部肿瘤患者的粗分类结果全部正确,细分类结果正确率为100%.结论:将等高轮廓线区域技术与Adaboost算法相结合,融合多个特征构建分类器提取并识别肺部肿瘤区域的方法能有效改善弱分类器的过拟合现象,有效的提高弱分类器的准确率,该算法实现了从PET-CT影像到诊断结果的自动化,为临床医生提供更清晰的诊断结果,极大提高临床工作效率.%Objective:To propose and design an automated process for localization of lesion region of lung and for assisted judgment of lesion sites by using PET-CT images, and to evaluate and analyze the whole algorithm flow so as to increase efficiency of clinical work.Methods: PET-CT images of 20 patients with lung tumor were selected and series of image processing methods including transforming of image scale were used to remove the bed of CT images. The contour line of equal altitude was used to extract region of sample in the image of PET-CT, and the region of sample was classified as category depended on pre-marked information of lesion region, and then the future of image in each region of sample was extracted. The Adaboost algorithm was applied to train and establish corresponding classification model. Finally, the classification model that has been trained was used to examine the test set, and the accuracy rate of integrated classification model consisted of weak classifiers was compared. Besides, the detectable rate, false detecting rate, ROC curve of interesting and the correct rate of the classification for cases were used to evaluate the results of classification.Results: There were 125088 samples were produced after the PET-CT images of 20 patients were pre- processed, and the positive samples and negative samples were 22720 and 102368, respectively. The correct rate of classification for sample of strong classifier, that was trained by using equal altitude contour line to classify region and using Adaboost. M2 algorithm to fuse with multi future, was around 97%. The results of rough classification of 20 patients with lung tumor were correct, and the correct rate of results of fine classification was 100%.Conclusion: The new method that combines the region technique of equal altitude contour line with Adaboost algorithm and that fuses multi futures to establish classifier and identify region of lunge tumor can efficiently increase the accurate rate of weak classifier. This method realizes the automation from PET-CT images to diagnosis results and provides clearer diagnosis results for clinicians, and increase the efficiency for clinical work.
展开▼
