目的 基于术前磁共振表观扩散系数(apparent diffusion coefficient,ADC)图建立预测成人颅内较低级别胶质瘤(lower-grade gliomas,LGG)1p/19q分子特征的影像组学模型并验证模型效能.材料与方法 回顾性分析于我院2017年1月至2021年12月之间术后病理证实、磁共振数据完整的LGG(WHO 2～3级)患者146例,其中1p/19q联合缺失(1p/19q co-deleted,1p/19q-Codel)68例,1p/19q未联合缺失(1p/19q non-codeleted,1p/19q-Noncodel)78例.按照7∶3的比例,采用完全随机法分为训练集与验证集.图像分割由一位影像医师使用ITK-SNAP软件独立进行,随后选取30例患者图像在医师间进行分割,用于评估所提取特征的稳定性.感兴趣容积(volume of interest,VOI)定义为液体衰减反转恢复(fluid-attenuated inversion-recovery,FLAIR)序列图像中除外明显囊变坏死的异常区域.将FLAIR图像上提取VOI复制至配准后的ADC图上,随后使用Python软件进行影像组学特征提取,并保留稳定性较好的特征进行Z-score标准化.Pearson或Spearman相关性分析以及最小绝对收缩与选择算子(least absolute shrinkage and selection operator,LASSO)分析用于特征选择.利用筛选出的组学特征建立影像组学评分(radiomics score,Rad-score)模型,采用受试者工作特征(receiver operating characteristic,ROC)曲线评估Rad-score模型的效能,并在验证集内进行验证.结果 146例LGG患者按7∶3的比例随机分为训练集(n=102)和验证集(n=44),两组之间患者的临床特征方面差异无统计学意义(P>0.05).通过组内和组间一致性分析、Pearson或Spearman相关性分析以及LASSO分析筛选出15个非零系数特征,并构建Rad-score模型.在训练集与验证集内,1p/19q-Codel组与1p/19q-Noncodel组在Rad-score上均存在差异(P<0.001).同时,Rad-score模型在训练集及验证集中均显示出良好的预测性能,在训练集中ROC曲线下面积(area under the curve,AUC)值为0.896,准确率85.29%,敏感度87.72%,特异度82.22%;验证集的AUC值0.778,准确率77.27%,敏感度71.43%,特异度82.61%.结论 基于术前ADC图建立的影像组学模型可以无创性预测成人颅内LGG的1p/19q分子特征.
ADC radiomics model in predicting 1p/19q molecular features of lower-grade gliomas
Objective:To establish and validate a radiomics model to predict 1p/19q molecular feature of adult intracranial lower-grade gliomas(LGG)based on preoperative magnetic resonance apparent diffusion coefficient(ADC)map.Materials and Methods:A total of 146 adult intracranial LGG(WHO grade 2-3)patients confirmed by postoperative pathology in our hospital from January 2017 to December 2021 with complete magnetic resonance data were retrospectively analyzed,including 68 cases with 1p/19q co-deleted(1p/19q-Codel)and 78 cases with 1p/19q non-codeleted(1p/19q-Noncodel).A completely random method was used to divide the training and validation sets in a 7:3 ratio.Image segmentation was performed independently by a radiologist using ITK-SNAP software,and 30 patient images were then segmented between radiologists to evaluate the stability of the extracted features.The volume of interest(VOI)was defined as the abnormal area in FLAIR,excluding obvious cystic and necrosis.The VOI extracted from the FLAIR image was copied to the registered ADC map,and then the radiomics features were extracted using Python software,and the features with good stability were retained for Z-score standardization.Pearson or Spearman correlation analysis and least absolute shrinkage and selection operator(LASSO)analysis were used for feature selection.The radiomics score(Rad-score)model was built using the selected radiomics features.The performance of the Rad-score model was evaluated using receiver operating characteristic(ROC)curves and validated within validation sets.Results:One hundred and forty six LGG patients were randomly divided into the training set(n=102)and the validation set(n=44)in a 7∶3 ratio.There was no statistical difference in clinical features between the two sets(P>0.05).Fifteen non-zero coefficient features were selected by intra-rater and inter-rater correlation coefficients,Pearson or Spearman correlation analysis and LASSO analysis,and the Rad-score model was constructed.There were significant differences in Rad-score between the 1p/19q-Codel and the 1p/19q-Noncodel in both the training and validation sets(P<0.001).At the same time,the Rad-score model showed good predictive performance in both the training and validation sets,with an area under the curve(AUC)value of 0.896 in the training set,the accuracy was 85.29%,the sensitivity was 87.72%and the specificity was 82.22%.The AUC value of the validation set was 0.778,the accuracy was 77.27%,the sensitivity was 71.43%and the specificity was 82.61%.Conclusions:The radiomics model based on the preoperative ADC map can noninvasively predict the 1p/19q molecular features in adult intracranial LGG.
国家科技期刊平台
NSTL
万方数据
