基于二维超声的病灶边缘区影像组学特征对三阴性乳腺癌术后复发的鉴别诊断

Differential diagnosis of postoperative recurrence of triple-negative breast cancer based on two-dimensional ultrasound radiomics features of margin area of the lesion

目的探讨术前边缘区影像组学分析在三阴性乳腺癌(TNBC)术后复发的鉴别诊断价值。方法回顾性收集2016年1月至2022年6月在海安市人民医院进行TNBC手术切除的139例患者资料,并依据随访结果分为复发组和未复发组。所有患者均接受乳腺二维超声(2DUS)检查。使用Image J基于高频灰阶声像图勾勒病灶边缘区,并提取感兴趣区(ROI)。基于Pyradiomics提取影像组学特征,并运用组内相关系数(ICC)、单因素分析及最小绝对值收缩和选择算子(LASSO)和自相关矩阵进行特征降维。采用Pearson相关分析影像组学特征与Ki-67指数及复发时间的关系。采用多元logistic回归构建影像组学模型,采用ROC曲线、校准曲线及决策曲线评价影像组学模型预测效能。结果排除11例失访者,最终纳入128例患者。按照7∶3比例分为训练集90例和测试集38例。训练集边缘区影像组学特征降维后最终得到5个关键特征(X1:original_glszm_GrayLevelNonUniformity,X2:original_glcm_Contrast,X3:wavelet-HLL_gldm_LowGrayLevelEmphasis,X4:wavelet-HHH_glszm_GrayLevelNonUniformity,X5:wavelet-HHH_glrlm_GrayLevelNonUniformityNormalized)。相关性分析证实X1、X2、X4与Ki-67指数均呈正相关,与复发时间均呈负相关(均P<0.05);X3和X5与Ki-67指数均呈负相关,与复发时间均呈正相关(均P<0.05)。多元logistic回归模型显示X2(OR=1.126,95%CI:1.086~1.165)、X3(OR=1.100,95%CI:1.056~1.143)和X5(OR=1.142,95%CI:1.109~1.172)是TNBC是否复发的独立影响因素。影像组学模型在训练集AUC为0.892(95%CI:0.828~0.962),验证集AUC为0.873(95%CI:0.809~0.943)。校准曲线及决策曲线显示影像组学模型具有良好的校准度和临床实用性。结论病灶边缘区影像组学特征分析具有良好的鉴别诊断效能,可有效预测TNBC术后复发风险。

Objective To investigate the value of preoperative marginal area radiomics analysis in the differential diagnosis of postoperative recurrence of triple-negative breast cancer(TNBC).Methods This study retrospectively collected data from 139 patients who underwent TNBC surgical resection at Haian People's Hospital from January 2016 to June 2022.Based on follow-up results,the patients were divided into recurrence and non-recurrence groups.All patients underwent breast two-dimensional ultrasound(2DUS)examination.By using Image J,the marginal area of the lesions was delineated on high-frequency grayscale images to extract the region of interest(ROI).Radiomics features were extracted using Pyradiomics.Intraclass correlation coefficient(ICC),univariate analysis,least absolute shrinkage and selection operator(LASSO)algorithm,and autocorrelation matrix were used for feature dimensionality reduction.Pearson correlation analysis was conducted to examine the relationship among radiomics features,Ki-67 index,and recurrence time.A logistic regression model was constructed based on radiomics features,and the model's differential diagnostic performance was evaluated using receiver operating characteristic(ROC)curves,calibration curves,and decision curves.Results After excluding 11 patients lost to follow-up,a total of 128 patients were included.Divided into a traning set of 90 cases and a testing set of 38 cases in a 7∶3 ratio.Comparison between groups showed that five key features(X1:original_glszm_GrayLevelNonUniformity,X2:original_glcm_Contrast,X3:wavelet-HLL_gldm_LowGrayLevelEmphasis,X4:wavelet-HHH_glszm_GrayLevelNon Unifor-mity,X5:wavelet-HHH_glrlm_Gray LevelNonUniformityNormalized)had statistically significant differences(P<0.05)after dimensionality reduction of marginal area radiomics features in the training set.Correlation analysis confirmed that X1,X2,and X4 were positively correlated with the Ki-67 index,and negatively correlated with recurrence time(all P<0.05),while X3 and X5 were negatively correlated with the Ki-67 index,and positively correlated with recurrence time(all P<0.05).Logistic regression model indicated that X2(OR=1.126,95%CI:1.086-1.165),X3(OR=1.100,95%CI:1.056-1.143),and X5(OR=1.142,95%CI:1.109-1.172)are independent influencing factors for TNBC recurrence.The model achieved an AUC of 0.892(95%CI:0.828-0.962)in the training set and an AUC of 0.873(95%CI:0.809-0.943)in the validation set.The calibration and decision curves demonstrated that the model has good calibration degree and clinical applicability.Conclusion Radiomics feature analysis on marginal area of the lesion has good differential diagnostic performance,and can effectively predict the risk of postoperative recurrence of TNBC.