| 陶响,刘梦玥,陈芳,庄英,顾清华.基于多模态MRI影像组学及SHAP分析预测直肠癌脉管侵犯[J].中国医学影像技术,2026,42(3):372~378 |
| 基于多模态MRI影像组学及SHAP分析预测直肠癌脉管侵犯 |
| Multimodal MRI radiomics and SHAP analysis for predicting lymphovascular invasion of rectal cancer |
| 投稿时间：2025-06-24 修订日期：2025-11-29 |
| DOI：10.13929/j.issn.1003-3289.2026.03.011 |
| 中文关键词: 直肠肿瘤 多模态成像 脉管侵犯 SHAP |
| 英文关键词:rectal neoplasms multimodal imaging lymphovascular invasion SHAP |
| 基金项目: |
|
| 摘要点击次数: 97 |
| 全文下载次数: 48 |
| 中文摘要: |
| 目的 观察多模态MRI影像组学及SHAP分析预测直肠癌(RC)脉管侵犯(LVI)的价值。方法 回顾性收集接受直肠MR检查的157例RC,按7∶3随机划分训练集(n=109)与测试集(n=48);分别于斜横轴位高分辨率T2WI及对比增强T1WI(T1C)中逐层分割病灶,提取病灶影像组学特征并建立T2WI、T1C及T1C-T2WI模型,以其中最佳者联合临床特征构建联合模型;绘制受试者工作特征(ROC)曲线,计算曲线下面积(AUC)评估模型效能,并以SHAP分析评价模型预测的可解释性。结果 MR壁外血管侵犯(EMVI)为临床独立预测因子,以之构建临床模型。分别于T2WI、T1C及T1C-T2WI中选出15、9及17个特征用于构建相应模型。T1C-T2WI模型为最佳影像组学模型,联合MR EMVI建立的联合模型在训练集和测试集的AUC分别为0.924和0.899,高于临床模型及T1C-T2WI模型,且预测概率和临床净收益均良好。SHAP分析揭示了影响多模态模型预测的关键变量。结论 基于多模态影像组学联合临床特征能有效预测RC LVI状态;利用SHAP可很好地解释特征对于模型预测概率的影响。 |
| 英文摘要: |
| Objective To explore the value of multimodal MRI radiomics and SHAP analysis for predicting lymphovascular invasion (LVI) of rectal cancer (RC). Methods Totally 157 RC patients who underwent rectal MR examination were retrospectively included and randomly divided into training set (n=109) and test set (n=48) at a ratio of 7∶3. ROI of lesions in oblique and transverse high-resolution T2WI and contrast-enhanced T1WI (T1C) were segmented layer by layer, radiomics features were extracted. T2WI, T1C and T1C-T2WI models were established, and the best radiomics model was selected to construct a combined model with clinical features. Receiver operating characteristic (ROC) curves were drawn, and the area under the curve (AUC) was calculated to evaluate efficacy of these models. SHAP analysis was then used to evaluate the interpretability of prediction of the models. Results MR extramural vascular invasion (EMVI) was identified as an independent clinical predictor, and a clinical model was constructed. Fifteen, 9 and 17 features were selected from T2WI, T1C and T1C-T2WI, respectively, and relative models were constructed. T1C-T2WI model was the best radiomics model. A combined model was established based on T1C-T2WI model combining with MR EMVI, and AUC of the combined model in training and test sets was 0.924 and 0.899, respectively, higher than that of clinical model and T1C-T2WI model, having good predicting probabilities and good clinical net returns. SHAP analysis revealed the key variables that affect the prediction of multimodal models. Conclusion Multimodal radiomics combined with clinical features could effectively predict LVI status of RC. SHAP could well explain the influence of features on model's prediction probability. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
