Diagnostic value of MR T1WI imaging omics in tuberculous spondylitis and brucella spondylitis
-
摘要:
目的 探讨MR T1WI影像组学的机器学习在结核性脊柱炎（TS）与布鲁菌性脊柱炎（BS）的鉴别诊断价值。 方法 回顾性收集我院经临床和实验室确诊的77例TS患者与34例BS患者的临床资料。按0.8:0.2的比例将病例分为训练集（n=88）和验证集（n=23）。首先由2位经验丰富的影像诊断医生对MR T1WI勾画感兴趣区域并提取MR T1WI影像组学特征，然后采用方差选择法、单变量特征选择法、最小绝对收缩与选择算子算法对组学特征信息进行选择和降维处理，挑选8项最具有特征的值。应用极限梯度增强法、Logistic回归法、支持向量机法、K邻近法4种机器学习算法构建MR T1WI影像组学在辨别确诊TS与BS模型。采用ROC曲线及曲线下面积（AUC）评判模型的确诊效能。 结果 极限梯度增强模型的AUC为0.80（95% CI：0.59~1.00），准确度为0.83；Logistic回归模型的AUC为0.85（95% CI：0.65~1.00），准确度为0.78；支持向量机模型的AUC为0.79（95% CI：0.62~0.97），准确度为0.78；K邻近法模型的AUC为0.75（95% CI：0.53~0.98），准确度为0.83。验证集中，Logistic回归模型的确诊效能最高。 结论 MR T1WI影像组学的机器学习在鉴别TS与BS的诊断效能较高，Logistic回归模型表现出更优的诊断效能。 Abstract:Objective To investigate the value of machine learning of MR T1WI radiomics in distinguishing tuberculous spondylitis (TS) from brucellar spondylitis (BS). Methods The clinical data of 77 TS patients and 34 BS patients diagnosed in our hospital were retrospectively collected. The patients were divided into training set (n=88) and verification set (n=23) according to the ratio of 0.8:0.2. First, two experienced diagnostic radiologists delineated the ROI on MR T1WI and extracted their imaging features. Then, we used the variance selection method, univariate feature selection method, LASSO for the feature information of radiomics to select and dimensionally reduce, and selected the 8 most characteristic values. Four machine learning algorithms, including XGBoost method, Logistic regression method, support vector machine method, and K-nearest neighbor method, were used to construct the MR T1WI imaging model in distinguishing TS from BS. The ROC curve and AUC were used to evaluate the diagnostic the effectiveness of these models. Results The AUC of XGBoost model was 0.80 (95% CI: 0.59-1.00), and the accuracy was 0.83. The AUC of Logistic regression model was 0.85 (95% CI: 0.65-1.00) and the accuracy was 0.78. The AUC of support vector machine model was 0.79 (95% CI: 0.62- 0.97), and the accuracy was 0.78. The AUC of K-nearest neighbor model was 0.75 (95% CI: 0.53-0.98), and the accuracy was 0.83. In the validation, Logistic regression model was the highest diagnostic efficiency in four machine learning algorithms. Conclusion The machine learning of MR T1WI radiomics has a good diagnostic efficiency in differentiating TS from BS. Logistic regression model is the best diagnostic efficiency in four machine learning algorithms. -
图 1 MR T1WI上手动勾画感兴趣区前后对比
Figure 1. Comparison before and after manual delineation of ROI on MR T1WI.
A: Before the outline; B: After the outline.
图 2 在特征选择上的套索算法
Figure 2. LASSO algorithm on feature selection.
A: MES path; B: LASSO path. Using Lasso model, 8 features which are correspond to the optimal alpha value were selected.
图 3 LASSO降维后通过模型选择筛选出的特征及其权重
Figure 3. Features and their weights filtered through model selection after LASSO dimensionality reduction.
图 4 4种机器学习算法预测模型在训练集中的ROC曲线
Figure 4. ROC curve of four machine learning algorithm prediction models in training set.
A: LR; B: SVM; C: KNN; D: XGBoost.
图 5 4种机器学习算法预测模型在验证集中的ROC曲线
Figure 5. ROC curve of four machine learning algorithm prediction models in validation set.
A: LR; B: SVM; C: KNN; D: XGBoost.
表 1 训练集与验证集患者性别、年龄比较
Table 1. Comparison of gender and age of patients in training set and verification set
Project Training set（n=88） Verification set（n=23） χ2/t P Gender（n） 0.589 0.443 Male 66 19 Female 22 4 Age（years, Mean±SD） 50.45±11.40 54.83±8.59 1.713 0.089 
下载: 导出CSV
表 2 4种T1WI机器学习算法预测模型在训练集中的诊断效能
Table 2. Diagnostic effectiveness of four T1WI machine learning algorithm prediction models in training set
Classifiers AUC 95% CI Precision LR 0.88 0.79-0.98 0.83 SVM 0.92 0.84-1.00 0.83 KNN 0.91 0.85-0.98 0.78 XGBoost 0.99 0.97-1.00 0.97 LR: Logistic regression; SVM: Support vector machine; KNN: K-nearest neighbor; XGBoost: Extreme gradient boosting.
下载: 导出CSV
表 3 4种T1WI机器学习算法预测模型在验证集中的诊断效能
Table 3. Diagnostic effectiveness of four T1WI machine learning algorithm prediction models in validation set
Classifiers AUC 95% CI Precision LR 0.85 0.84-1.00 0.78 SVM 0.79 0.87-1.00 0.78 KNN 0.75 0.82-1.00 0.83 XGBoost 0.80 0.62-1.00 0.83
下载: 导出CSV
-
[1] 李立新, 阮锦荣, 孙莉. MRI与CT诊断及鉴别脊柱结核与脊柱肿瘤的价值观察[J]. 中国CT和MRI杂志, 2021, 19(8): 175-7. https://www.cnki.com.cn/Article/CJFDTOTAL-CTMR202108057.htm [2] Liao JC, Lai PL, Chen LH, et al. Surgical outcomes of infectious spondylitis after vertebroplasty, and comparisons between pyogenic and tuberculosis[J]. BMC Infect Dis, 2018, 18(1): 555. doi: 10.1186/s12879-018-3486-x [3] 帕哈提·吐逊江, 杨来红, 何雄, 等. 影像组学在脊柱疾病中的应用[J]. 磁共振成像, 2022, 13(5): 162-6. https://www.cnki.com.cn/Article/CJFDTOTAL-CGZC202205036.htm [4] Joshi RS, Lau D, Ames CP. Artificial intelligence for adult spinal deformity: current state and future directions[J]. Spine J, 2021, 21 (10): 1626-34. doi: 10.1016/j.spinee.2021.04.019 [5] Yu JH, Shi ZF, Lian YX, et al. Noninvasive IDH1 mutation estimation based on a quantitative radiomics approach for grade Ⅱ glioma[J]. Eur Radiol, 2017, 27(8): 3509-22. doi: 10.1007/s00330-016-4653-3 [6] 黄冠, 尹芳艳, 李小雪, 等. 影像组学研究方法进展[J]. 实用放射学杂志, 2019, 35(2): 308-11. https://www.cnki.com.cn/Article/CJFDTOTAL-FSXS202011032.htm [7] 蒋西然, 蒋韬, 孙嘉瑶, 等. 深度学习人工智能技术在医学影像辅助分析中的应用[J]. 中国医疗设备, 2021, 36(6): 164-71. https://www.cnki.com.cn/Article/CJFDTOTAL-YLSX202106040.htm [8] Wang K, Qiao Z, Zhao XB, et al. Individualized discrimination of tumor recurrence from radiation necrosis in glioma patients using an integrated radiomics-based model[J]. Eur J Nucl Med Mol Imaging, 2020, 47(6): 1400-11. doi: 10.1007/s00259-019-04604-0 [9] Peng ZY, Wang YM, Wang YX, et al. Application of radiomics and machine learning in head and neck cancers[J]. Int J Biol Sci, 2021, 17(2): 475-86. doi: 10.7150/ijbs.55716 [10] 吴林永, 赵羽佳, 林鹏, 等. 基于人工智能术前预测乳腺导管内癌微浸润的价值[J]. 中国医学影像学杂志, 2021, 29(1): 29-34. https://www.cnki.com.cn/Article/CJFDTOTAL-ZYYZ202101011.htm [11] 蒋忠军, 蒋小龙, 文建荣, 等. 基于影像组学分析的良恶性乳腺肿瘤临床诊断与预测模型研究[J]. 实用放射学杂志, 2022, 38(2): 236-9. [12] 阴彦林, 杨新明, 田野, 等. MRI和PET/CT在感染性脊柱炎鉴别诊断中的应用价值[J]. 实用放射学杂志, 2021, 37(10): 1673-6, 1684. [13] Granzier RWY, Ibrahim A, Primakov SP, et al. MRI-based radiomics analysis for the pretreatment prediction of pathologic complete tumor response to neoadjuvant systemic therapy in breast cancer patients: a multicenter study[J]. Cancers, 2021, 13(10): 2447. [14] Yamamoto T, Kimura H, Hayashi K, et al. Pseudo- continuous arterial spin labeling MR images in Warthin tumors and pleomorphic adenomas of the parotid gland: qualitative and quantitative analyses and their correlation with histopathologic and DWI and dynamic contrast enhanced MRI findings[J]. Neuroradiology, 2018, 60(8): 803-12. [15] Liu JF, Wang CJ, Guo W, et al. A preliminary study using spinal MRI-based radiomics to predict high-risk cytogenetic abnormalities in multiple myeloma[J]. Radiol med, 2021, 126(9): 1226-35. [16] Li, He, MM, et al. Radiomics based on lumbar spine magnetic resonance imaging to detect osteoporosis[J]. Acad Radiol, 2021, 28 (6): e165-71. [17] Daimiel Naranjo I, Gibbs P, Reiner JS, et al. Radiomics and machine learning with multiparametric breast MRI for improved diagnostic accuracy in breast cancer diagnosis[J]. Diagnostics (Basel), 2021, 11(6): 919. -
点击查看大图
计量
- 文章访问数: 110
- HTML全文浏览量: 119
- PDF下载量: 10
- 被引次数: 0
