【摘要】 目的 构建基于MRI影像组学特征的机器学习模型,探讨其在鉴别诊断鞍区病变中的应用价值。方法 2016年1月—2021年2月四川大学华西医院诊治鞍区病变患者259例,其中颅咽管瘤（CR）81例、垂体腺瘤（PA）63例、Rathke’s囊肿（RCC）61例、鞍结节脑膜瘤（TSM）54例,均行常规MRI检查（T₁WI、T₂WI、T₁WI增强序列）,从每个MRI序列原始图像中提取影像组学特征。将所有患者随机分为4个训练集和1个验证集,基于提取的影像组学特征、采用支持向量机（SVM）、逻辑回归（LR）、极端梯度增强（XGBoost）机器学习算法在训练集中构建诊断鞍区病变的模型。在验证集中采用混淆矩阵评价3种机器学习模型在不同MRI序列中诊断鞍区病变的平衡准确率;绘制宏平均ROC曲线,评估3种机器学习模型在不同MRI序列中诊断鞍区病变的整体效能。以组织病理检查结果为金标准,计算XGBoost模型在不同MRI序列中诊断4种鞍区病变的灵敏度、特异度和准确率;采用F-score、协方差分析评价XGBoost模型在T₁WI增强序列中诊断鞍区病变的特征重要性。结果 XGBoost模型在T₁WI、T₂WI、T₁WI增强序列中诊断鞍区病变的平衡准确率（62.3%、75.4%、82.8%）均高于SVM模型（61.2%、71.8%、76.7%）、LR模型（60.2%、72.1%、75.4%）。SVM、LR、XGBoost模型在T₁WI序列中诊断鞍区病变的平均AUC分别为0.846（95%CI:0.827～0.865,P<0.001）、0.820（95%CI:0.789～0.851,P<0.001）、0.852（95%CI:0.816～0.889,P<0.001）,在T₂WI序列中诊断鞍区病变的平均AUC分别为0.925（95%CI:0.906～0.944,P<0.001）、0.922（95%CI:0.884～0.960,P<0.001）、0.931（95%CI:0.898～0.965,P<0.001）,在T₁WI增强序列中诊断鞍区病变的平均AUC分别为0.938（95%CI:0.915～0.961,P<0.001）、0.929（95%CI:0.907～0.950,P<0.001）、0.956（95%CI:0.942～0.970,P<0.001）。XGBoost模型在T₁WI增强序列中诊断4种鞍区病变的灵敏度、特异度比较差异均有统计学意义（χ²=4.597,P=0.032;χ²=4.020,P=0.045）,准确率比较差异无统计学意义（χ²=1.538,P=0.215）;XGBoost模型在T₁WI、T₂WI序列中诊断4种鞍区病变的灵敏度、特异度、准确率比较差异均无统计学意义（χ²=0.122～3.807,P均>0.05）;XGBoost模型在T₁WI增强序列中诊断TSM的灵敏度（86.7%）高于CR（82.4%）、PA（78.4%）、RCC（82.8%）（P<0.05）,诊断CR的特异度（97.5%）高于PA（93.2%）、RCC（93.4%）、TSM（93.8%）（P<0.05）。XGBoost模型在T₁WI增强序列中诊断鞍区病变的前3位重要特征分别为original＿firstorder＿Variance、wavelet-HHL＿firstorder＿Mean、wavelet-HHL＿firstorder＿Skewness,F-score分别为58、57、39;协方差热图显示,T₁WI增强序列中鞍区病变的影像组学特征与对应个体间的关联模式存在明显差异,区分CR、PA、RCC、TSM的性能较好。结论 基于影像组学特征构建的XGBoost模型在T₁WI、T₂WI、T₁WI增强序列中诊断鞍区病变（CR、PA、RCC、TSM）的效能均优于SVM和LR模型,其中在T₁WI增强序列中区分4种鞍区病变的效能最佳。更多还原

【Abstract】 Objective To construct machine learning models based on MRI radiomic features and explore their value in the differential diagnosis of sellar region lesions.Methods A total of 259 patients with sellar region lesions treated in West China Hospital of Sichuan University from January 2016 to February 2021 were included,among whom there were 81 cases of craniopharyngioma（CR）,63 cases of pituitary adenoma（PA）,61 cases of Rathke’s cleft cyst（RCC）,and 54 cases of tuberculum sellar meningioma（TSM）.All patients underwent conventional MRI examinations including T₁WI,T₂WI,and contrast-enhanced T₁WI sequences.Radiomic features were extracted from the original images of each MRI sequence.The patients were randomly divided into 4 training sets and 1 validation set.Based on the extracted radiomic features,diagnostic models for sellar region lesions were constructed in the training sets using support vector machine（SVM）,logistic regression（LR）,and extreme gradient boosting（XGBoost）machine learning algorithms.In the validation set,the balanced accuracy of the three machine learning models in diagnosing sellar region lesions across different MRI sequences was evaluated using a confusion matrix.Macro-averaged ROC curves were plotted to assess the overall performance of the three machine learning models in diagnosing sellar region lesions across different MRI sequences.Using histopathological examination as the gold standard,the sensitivity,specificity,and accuracy of the XGBoost model in diagnosing four types of sellar region lesions on different MRI sequences were calculated.The F-score and covariance analysis were used to evaluate the feature importance of the XGBoost model in diagnosing sellar region lesions on contrast-enhanced T₁WI sequence.Results The balanced accuracies of the XGBoost model in diagnosing sellar region lesions on T₁WI,T₂WI,and contrast-enhanced T₁WI sequences（62.3%,75.4%,82.8%）were higher than those of the SVM model（61.2%,71.8%,76.7%）and the LR model（60.2%,72.1%,75.4%）.The average AUCs of the SVM,LR,and XGBoost models for diagnosing sellar region lesions on T₁WI sequence were 0.846（95%CI:0.827-0.865,P<0.001）,0.820（95%CI:0.789-0.851,P<0.001）,and 0.852（95%CI:0.816-0.889,P<0.001）,respectively;on T₂WI sequence,they were 0.925（95%CI:0.906-0.944,P<0.001）,0.922（95%CI:0.884-0.960,P<0.001）,and 0.931（95%CI:0.898-0.965,P<0.001）,respectively;on contrast-enhanced T₁WI sequence,they were 0.938（95%CI:0.915-0.961,P<0.001）,0.929（95%CI:0.907-0.950,P<0.001）and 0.956（95%CI:0.942-0.970,P<0.001）,respectively.For the XGBoost model on contrast-enhanced T₁WI sequence,there were statistically significant differences in sensitivity and specificity for diagnosing the four types of sellar region lesions（χ²=4.597,P=0.032;χ²=4.020,P=0.045）,but no statistically significant difference in accuracy （χ²=1.538,P=0.215）.For the XGBoost model on T₁WI and T₂WI sequences,there were no statistically significant differences in sensitivity,specificity,or accuracy for diagnosing the four types of sellar region lesions （χ²=0.122-3.807,all P values>0.05）.On contrast-enhanced T₁WI sequence,the XGBoost model showed higher sensitivity for diagnosing TSM（86.7%）compared with CR（82.4%）,PA（78.4%）,and RCC（82.8%）（P<0.05）,and higher specificity for diagnosing CR（97.5%）compared with PA（93.2%）,RCC（93.4%）,and TSM（93.8%）（P<0.05）.The top three important features of the XGBoost model for diagnosing sellar region lesions on contrast-enhanced T₁WI sequence were original＿firstorder＿Variance,wavelet-HHL＿firstorder＿Mean,and wavelet-HHL＿firstorder＿Skewness,with F-scores of 58,57,and 39,respectively.The covariance heatmap revealed distinct association patterns between radiomic features of sellar region lesions and corresponding individuals on contrast-enhanced T₁WI sequence,demonstrating good performance in distinguishing CR,PA,RCC,and TSM.Conclusion The XGBoost model constructed based on radiomic features outperforms the SVM and LR models in diagnosing sellar region lesions（CR,PA,RCC,TSM）on T₁WI,T₂WI,and contrast-enhanced T₁WI sequences,with the best performance in differentiating the four types of sellar region lesions observed on contrast-enhanced T₁WI sequence.更多还原