摘要
目的探讨数据挖掘技术评价MRI乳腺影像报告和数据系统（BI—RADS—MRI）分析乳腺非肿块样强化病灶的价值。方法对55例行乳腺MR检查，且MRI增强表现为非肿块样强化病灶的患者，应用决策树和Logistic回归2种分类模型，对55个病灶进行分析。再将资料进行“10-折交叉验证”。结果55个病灶包括28个恶性肿瘤和27个良性病灶。簇状小环样和块状强化多出现在恶性病灶中[分别为12和d个；决策树模型中占8d．2％（16／19），Logistic模型中偏回归系数分别为2．128和1．723]。均匀、丛状、网状和线样导管样强化更易出现在良性病灶中[分别为4、9、1和7个；决策树模型中占72．4％（21／29），Logistic模型中偏回归系数分别为均匀0．357，丛状1．861，网状18．870]。10-折交叉验证评估2种模型Logistic回归的准确度57．0％，敏感度43．3％，特异度71．7％；决策树的准确度69．3％，敏感度66．7％，特异度71．7％。结论基于BI—RADS—MRI的非肿块样强化描述对该类病变作出正确诊断的效能不高，进一步对各种征象作出补充非常必要。
Objective To evaluate the diagnostic values of the breast imaging reporting and data system-MRI （BI-RADS-MRI）description about non-masslike enhancement by data mining. Methods Fiftyfive patients with non-masslike enhancement lesions showed on breast contrast-enhanced MRI were evaluated using two data mining algorithms （Logistic regression and decision tree ） and 10-fold cross-validation methods. Results There were 28 malignant and 27 benign lesions. The most frequent findings of the malignant lesions were clustered ring enhancement and clumped enhancement [ 12 and 4 lesions, respectively; 84. 2% （16/19） in decision trees, partial regression coefficients in Logistic model were 2. 128 and 1.723, respectively], whereas homogenous, stippled, reticular internal and linear ductal enhancement were the most frequent findings in benign lesions [ 4,9,1 and 7 lesions, respectively; 72.4% （21/29） in decision tree, partial regression coefficients in Logistic model were 0. 357 （homogenous）, 1. 861 （stippled） and 18. 870（ reticular）, respectively]. 10-fold cross-validation indicated that decision tree （C5.0） achieved an accuracy of 69. 3% with a sensitivity of 66. 7% and a specificity of 71.7% in comparison to the Logistic regression model with an accuracy of 57. 0%, a sensitivity of 43.3% and a specificity of 71.7%. Conclusions The diagnosis efficacy of non-masslike enhancement interpretation according to BI-RADS-MRI is not high. It is very important to find more potential features of non-masslike enhancement to improve the diagnosis accuracy.
出处
《中华放射学杂志》
CAS
CSCD
北大核心
2009年第5期455-459,共5页
Chinese Journal of Radiology
关键词
乳腺肿瘤
数据显示
磁共振成像
交叉研究
Breast neoplasms
Data display
Magnetic resonance imaging
Cross-over studies
