超声BI-RADS术语特征用于量化BI-RADS 3~5评估分类的研究

Ultrasonic BI-RADS Terminology Features Used to Quantify BI-RADS 3~5 Evaluation Classification

目的 利用US-BI-RADS对乳腺病变进行规范化描述,采用决策树机器学习,为BI-RADS 3~5类提供一个量化指标。方法 对每个术语特征进行哑变量设置,首先进行单因素Logistic回归,差异有统计学意义的哑变量进入二元Logistic回归分析并进行赋值,将每个术语特征的分值相加得到一个总分值;以病理结果为依据,对总分值采用决策树机器学习并进行10折交叉验证,按5分类法将不同总分值分为BI-RADS 3~5类。结果 恶性病变患者年龄明显大于良性病变患者,差异有统计学意义(P<0.001)。单因素Logistic回归分析显示术语特征变量中不规则、不平行、模糊、成角、微小分叶、毛刺、声影、钙化(微钙化、粗钙化)、结构扭曲、水肿、内部血供以及腋窝淋巴结异常差异有统计学意义,二元Logistic回归分析,得出上述的变量的赋值情况,不规则2分、不平行1分、模糊1分、成角1分、微小分叶1分、毛刺1分、钙化(微钙化2分、粗钙化-1分)、结构扭曲1分、水肿1分、内部血供1分以及腋窝淋巴结异常1分,采用决策树机器学习10折交叉验证得到3类、4A类、4B类、4C类、5类对应的分值分别为≤0分、1~2分、3~5分、6~7分以及≥8分,验证后3~5分类恶性风险分别为0.43%、5.44%、34.92%、85.56%以及98.38%。AUC为0.951,当以≥5分为预测恶性的cut-off值时,其敏感度、特异度、准确度分别为87.6%、89.5%及89.9%。结论 超声BI-RADS术语特征及最终评估分类对乳腺病变规范化描述和临床处理建议有重要的参考价值,不同超声图像术语特征分值相加,分值越大,恶性的概率越高,量化指标使得BI-RADS评估分类具有更好的可行性。

Objective To use US-BI-RADS to describe standardized term features of breast lesions,and assign different term features.The sum of term feature scores adopts decision tree machine learning to provide a quantitative index for BI-RADS categories 3 to 5.Methods Dummy variables were set for each term feature.The set dummy variables were first subjected to uni-variant logistic regression.The dummy variables with statistically significant differences were entered into binary logistic regression analysis and assigned values.The scores of each term feature were added to obtain a total score.Based on the pathological results,malignant risk of different total scores was calculated,and total scores were studied by decision tree machine learning and 10 fold cross validation.Different total scores were divided into BI-RADS 3~5 categories according to 5 classification methods.Results Age of patients with malignant lesions was significantly older than that of patients with benign lesions with statistically significant difference(P<0.001).Uni-variant logistic regression analysis showed that there were statistically significant differences in terms of irregular,non parallel,fuzzy,angular,micro lobulation,burr,acoustic shadow,calcification(micro calcification,coarse calcification),structural distortion,edema,internal blood supply and axillary lymph node abnormalities among term characteristic variables.Binary logistic regression analysis to obtain assignment of the above variables,2 points for irregular,1 point for non parallel,1 point for fuzzy,1 point for angled,1 point for micro lobulation,1 point for burr,calcification(2 points for micro calcification,-1 point for coarse calcification),1 point for structural distortion,1 point for edema,1 point for internal blood supply and 1 point for axillary lymph node abnormality.According to the 10 fold cross validation of policy tree machine learning,the corresponding scores of category 3,category 4A,category 4B,category 4C and category 5 are respectively≤0 point,1~2 points,3~5 points,6~7 points and≥8 points.After verification,malignant risk of 3~5 classification was 0.43%,5.44%,34.92%,85.56%and 98.38%,respectively.The AUC was 0.951.When the cut-off value was≥5,sensitivity,specificity and accuracy were 87.6%,89.5%and 89.9%,respectively.Conclusion Ultrasonic BI-RADS terminology features and final evaluation and classification have important reference value for standardized description of breast lesions and clinical treatment suggestions.Different ultrasonic image terminology feature scores are added,and the higher the score,the higher the probability of malignancy.Quantitative indicators would make BI-RADS evaluation and classification more feasible.