基于MRI DWI的影像组学模型对骨肉瘤新辅助化疗疗效的评估价值

目的 探讨基于MRI弥散加权成像（DWI）的影像组学模型对骨肉瘤新辅助化疗疗效的评估价值。 方法 回顾性分析河北医科大学第三医院2015年6月至2017年11月经术后组织病理学检查结果证实，且在接受新辅助化疗前、后均行MRI检查的41例骨肉瘤患者[男性26例、女性15例，年龄（22.0±11.0）岁，范围11~49岁]的病历及影像资料。根据术后组织病理学检查结果，将肿瘤组织坏死率≥90%者纳入疗效好组，<90%者纳入疗效差组。分别于新辅助化疗前、化疗一期结束后5 d内和完成整个化疗后测量所有患者的表观扩散系数（ADC，分别记为ADC0、ADC1、ADC2，比较疗效好组和疗效差组ADC间的差异。在化疗一期结束后的DWI（b=1000 s/mm2）和ADC图像上手动勾画病灶的感兴趣区，提取影像组学特征，用随机分组法将数据按6∶4的比例分为训练集和测试集，采用SMOTE算法对训练集上的数据进行扩充，采用方差阈值、SelectKBest、最小绝对收缩和选择算子（LASSO）法进行影像组学特征筛选，采用逻辑回归分类器构建出影像组学模型。采用独立样本t检验或Wilcoxon秩和检验进行2组间比较。采用受试者工作特征（ROC）曲线评估传统影像学（ADC）及影像组学模型对骨肉瘤新辅助化疗疗效的预测效能。
结果 疗效好组10例、疗效差组31例。2组患者ADC0的差异无统计学意义[（0.95±0.05）×10−3 mm2/s对（1.05±0.05）×10−3 mm2/s，t=1.14，P>0.05）]；疗效好组的ADC1、ADC2高于疗效差组，且差异均有统计学意义[（1.44±0.10）×10−3 mm2/s对（1.10±0.06）×10−3 mm2/s，t=−2.92，P<0.05；1.68（1.55，1.85）×10−3 mm2/s对（1.33±0.06）×10−3 mm2/s，Z=−2.61，P<0.01]。ROC曲线分析结果显示，当ADC1≥1.34×10−3 mm2/s时，其对骨肉瘤新辅助化疗疗效评估的灵敏度为80%，特异度为81%，曲线下面积（AUC）为0.797（95%CI：0.629~0.965）；当ADC2≥1.51×10−3 mm2/s时，其对骨肉瘤新辅助化疗疗效评估的灵敏度为90%，特异度为71%，AUC为0.777（95%CI：0.588~0.967）。从化疗一期结束后的DWI和ADC图像中共提取出1409个影像组学特征，按6∶4的比例随机分为训练集和测试集[24（疗效好：6，疗效坏：18）∶17（疗效好：4，疗效坏：13）]，将训练集数据扩充为70（疗效好：20，疗效坏：50），经影像组学特征筛选后，最终得到5个最优影像组学特征，包括Interquartile Range、Skewness、Uniformity、Median、Maximum。采用逻辑回归分类器构建影像组学模型，训练集中该模型预测骨肉瘤新辅助化疗疗效的ROC曲线的AUC为0.881（95%CI：0.811~0.942），灵敏度为90%，特异度为74%；测试集中AUC为0.769（95%CI：0.515~0.933），灵敏度为75%，特异度为69%。 结论 基于MRI DWI的影像组学模型在评估骨肉瘤新辅助化疗疗效中的效能优于传统影像学（ADC），在临床应用中潜力较大。

Evaluation of the efficacy of neoadjuvant chemotherapy in osteosarcoma based on MRI DWI radiomics model

Objective To investigate the value of MRI DWI-based radiomics models for evaluating the treatment response in osteosarcoma after neoadjuvant chemoherapy. Methods A retrospective analysis was conducted on the medical records and imaging data of 41 patients with osteosarcoma (26 males and 15 females; aged (22.0±11.0) years, range of 11–49 years) who underwent MRI examinations before and after receiving neoadjuvant chemotherapy, and confirmed by postoperative histopathological examinations at the Third Hospital of Hebei Medical University from June 2015 to November 2017. In accordance with the postoperative histopathological examination results, patients with a tumor tissue necrosis rate of ≥90% were included in the good-efficacy group, and those with a necrosis rate of <90% were included in the poor-efficacy group. The apparent diffusion coefficient (ADC, denoted as ADC0, ADC1, and ADC2) were measured in all patients before neoadjuvant chemotherapy, within 5 days after the end of the first stage of chemotherapy, and after the completion of the entire chemotherapy. The differences in ADC were compared between the two groups. The region of interest of the lesion was manually delineated on DWI (b=1 000 s/mm2) and ADC images after the end of the first stage of chemotherapy, and the radiomics features were extracted. Data were divided into training set and validation set by using random grouping at 6∶4. The SMOTE algorithm was used to expand the data on the training set. The variance threshold, SelectKBest, and least absolute shrinkage and selection operator (LASSO) algorithm were used to screen the radiomics features. A radiomics model was constructed using a logistic regression classifier. Independent sample t-test or Wilcoxon rank sum test was used to compare the two groups. Receiver operating characteristic (ROC) curves were used to evaluate the predictive efficacy of traditional imaging (ADC) and radiomics models on the efficacy of neoadjuvant chemotherapy for osteosarcoma. Results A total of 10 and 31 cases were included in the good-efficacy and poor-efficacy groups, respectively. No statistically significant difference was found in the ADC0 value between the two groups ((0.95±0.05)×10−3 mm2/s vs. (1.05±0.05)×10−3 mm2/s, t=1.14, P>0.05)). The values of ADC1 and ADC2 in the good-efficacy group were higher than those in the poor-efficacy group, with statistical significance ((1.44±0.10)×10−3 mm2/s vs. (1.10±0.06)×10−3 mm2/s, t=−2.92, P<0.05; 1.68 (1.
55, 1.85)×10−3 mm2/s vs. (1.33±0.06)×10−3 mm2/s, Z=−2.61, P<0.01). ROC curve analysis showed that when ADC1 ≥1.34×10−3 mm2/s, the sensitivity for evaluating the efficacy of neoadjuvant chemotherapy in osteosarcoma was 80%, the specificity was 81%, and the area under the curve (AUC) was 0.797 (95%CI: 0.629–0.965). When ADC2 ≥1.51×10−3 mm2/s, the sensitivity for evaluating the efficacy of neoadjuvant chemotherapy in osteosarcoma was 90%, the specificity was 71%, and the AUC was 0.777 (95%CI: 0.588–0.967). A total of 1 409 radiomics features were extracted from the DWI and ADC images after the end of the first stage of chemotherapy. They were randomly divided into training set and validation set at a ratio of 6∶4 (24 (good efficacy: 6, poor efficacy: 18)∶17 (good efficacy: 4, poor efficacy: 13)). The training set data were expanded to 70 (good efficacy: 20, poor efficacy: 50). After the radiomics features were screened, five optimal radiomics features were ultimately obtained, including InterquartileRange, Skewness, Uniformity, Median, and Maximum. Logistic regression classifier was used to construct a radiomics model. The ROC curves showed that in the training set, the AUC of the model for predicting the efficacy of neoadjuvant chemotherapy in osteosarcoma was 0.881 (95%CI: 0.811–0.942), with sensitivity of 90% and specificity of 74%. Meanwhile, in the validation set, the AUC was 0.769 (95%CI: 0.515–0.933), with sensitivity of 75% and specificity of 69%. Conclusion The radiomics model based on MRI DWI outperforms the traditional imaging (ADC) in evaluating the efficacy of neoadjuvant chemotherapy for osteosarcoma, showing great potential in clinical applications.