| 杨晓楠,王得志,王成健,郝大鹏,徐文坚,崔久法.多参数MRI影像组学与深度学习模型鉴别良、恶性黏液样软组织肿瘤[J].中国医学影像技术,2024,40(7):1078~1082 |
| 多参数MRI影像组学与深度学习模型鉴别良、恶性黏液样软组织肿瘤 |
| Differentiating benign and malignant myxoid soft tissue tumors based on multiparametric MRI radiomics and deep learning models |
| 投稿时间：2024-01-05 修订日期：2024-04-03 |
| DOI：10.13929/j.issn.1003-3289.2024.07.024 |
| 中文关键词: 软组织肿瘤 磁共振成像 深度学习 影像组学 |
| 英文关键词:soft tissue neoplasms magnetic resonance imaging deep learning radiomics |
| 基金项目: |
|
| 摘要点击次数: 500 |
| 全文下载次数: 218 |
| 中文摘要: |
| 目的 观察基于多参数MRI构建的影像组学与深度学习(DL)模型鉴别良、恶性黏液样软组织肿瘤(MSTT)的价值。方法 回顾性纳入141例经病理证实的MSTT患者,以7 ∶ 3比例随机将其分为训练集(n=98,包括51例恶性及47例良性MSTT)及测试集(n=43,包括22例恶性及21例良性MSTT)。分别于训练集T1WI和脂肪抑制(FS)-T2WI中提取并遴选影像组学特征及DL特征,并以之构建鉴别良、恶性MSTT的影像组学模型及DL模型。绘制受试者工作特征(ROC)曲线、校准曲线及决策曲线,对比评估2个模型的区分度、校准度及净收益。结果 于训练集提取并筛选得到9个最佳影像组学特征用于构建鉴别良、恶性MSTT的影像组学模型,包括2个一阶特征、1个形态特征、3个灰度共生矩阵特征、1个灰度相关矩阵特征和2个灰度大小区域矩阵特征;以其中7个最佳DL特征构建DL模型。所获影像组学模型和DL模型鉴别测试集良、恶性MSTT的ROC曲线下面积分别为0.758及0.911,后者高于前者(P=0.017);2个模型均具有良好校准度;相比影像组学模型,DL模型在测试集的总体净收益更高。结论 基于MRI构建的DL模型鉴别良、恶性MSTT的效能较影像组学模型更好且净收益更高。 |
| 英文摘要: |
| Objective To observe the value of multiparametric MRI-based radiomics model and deep learning (DL) model for distinguishing benign and malignant myxoid soft tissue tumors (MSTT). Methods A total of 141 MSTT patients confirmed with pathology were retrospectively collected. The patients were randomly divided into training set (n=98, including 51 cases of malignant MSTT and 47 cases of benign MSTT) and test set (n=43, including 22 cases of malignant MSTT and 21 cases of benign MSTT) at the ratio of 7∶3. Based on T1WI and fat suppression (FS)-T2WI in training set, radiomics features and DL features were extracted and selected, then a radiomics model and a DL model were constructed, respectively. Receiver operating characteristic (ROC) curves, calibration curves and decision curves were drawn, and the discrimination, calibration and net benefit of these 2 models were compared. Results In training set, the radiomics model for differentiating benign and malignant MSTT was constructed according to 9 optimal radiomics features, including 2 first order features, 1 shape feature, 3 gray level co-occurrence matrix features, 1 gray level dependence matrix feature and 2 gray level size zone matrix features, while DL model was built based on 7 optimal DL features. In test set, the area under the ROC curve of radiomics model and DL model was 0.758 and 0.911, respectively, the latter was higher than the former (P=0.017). Both models had good calibration, and DL model had higher overall net benefit. Conclusion Compared with radiomics model, DL model based on MRI had better ability to differentiating benign and malignant MSTT, also higher overall net benefit. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
