目的 探讨基于增强CT影像组学模型在肾透明细胞癌（ccRCC）和肾嗜酸细胞腺瘤（RO）鉴别诊断中的价值。 方法 回顾性队列研究。纳入山西省肿瘤医院2013年6月—2022年7月收治的经病理确诊的肾肿瘤患者100例，其中男63例、女37例，年龄42~81（60.8±8.7）岁。100例中，ccRCC患者75例（ccRCC组），其中男52例、女23例，年龄42~79（59.9±8.5）岁；RO患者25例（RO组），其中男11例、女14例，年龄46~81（63.5±8.8）岁。将入组患者按照3∶7随机分为测试集（30例）和训练集（70例）。患者均行增强CT检查。从门静脉期CT图像中提取1 409个影像组学特征，使用Variance Threshold、SelectKBest以及最小绝对收缩与选择算子算法进行特征筛选，构建影像组学模型；绘制受试者操作特征曲线（ROC曲线），计算曲线下面积（AUC）。用灵敏度、特异度、精确度以及F1-score来评估不同影像组学模型诊断效果。 结果 ccRCC组和RO组患者的年龄比较，差异无统计学意义（t=1.82，P=0.072）；2组患者的性别比较，差异有统计学意义（χ 2=5.16，P=0.023）。1 409个影像组学特征中最终筛选出12个最具相关性的组学特征，建立5个影像组学模型用于鉴别ccRCC和RO，分别为支持向量机、逻辑回归（LR）模型、决策树模型、K-近邻模型、随机森林模型。结果表明，这5个模型在训练集和测试集中的AUC值分别为0.905［95%可信区间（CI） 0.826~0.984］、0.870（95%CI 0.742~0.996），0.910（95%CI 0.717~0.989）、0.853（95%CI 0.717~0.989），0.885（95%CI 0.787~0.983）、0.628（95%CI 0.353~0.903），0.925（95%CI 0.873~0.977）、0.638（95%CI 0.416~0.861），0.980（95%CI 0.954~1.000）、0.821（95%CI 0.673~0.968）。综合各项指标可见，在5个模型中LR模型性能最优，具有较理想的诊断效能。 结论 基于增强CT的影像组学模型，可以提高对ccRCC和RO鉴别诊断的准确性。 Objective This study aimed to explore the value of enhanced CT imaging model combined with machine learning in the differential diagnosis of renal clear cell carcinoma (ccRCC) and renal eosinophilic adenoma (RO). Methods Retropective cohort study was conduted. A total of 100 patients with renal cancer diagnosed pathologically admitted to Shanxi Cancer Hospital from June 2013 to July 2022 were included, including 63 males and 37 females, aged 42-81 (60.8±8.7) years. Of the 100 patients, 75 were pathologically confirmed as ccRCC patients, including 52 males and 23 females, aged 42-79 (59.9±8.5) years. There were 25 RO patients in the RO group, including 11 males and 14 females, aged 46-81 (63.5±8.8) years. The patients were classified into test cohort (30%) and training cohort (70%) with a random method, and all of them underwent enhanced CT examination. Then, 1 409 imaging features were extracted from portal venous CT images, and variance threshold, SelectKBest, and least absolute shrinkage and selection operator algorithms were used in feature extraction. We constructed an image omics model and drew the receiver operating characteristic curve. Sensitivity, specificity, accuracy, and F1-score were used in evaluating the performance of different imaging models. Results No statistically significant difference in age was found between the groups (t=-1.82, P=0.072). The gender showed statistically significant difference between the groups (χ2=5.16, P=0.023). The 12 most relevant features were selected from 1 409 image omics features, and five models were established to differentiate between ccRCC and RO: support vector machine, logistic regression (LR) model, decision tree model, K-nearest neighbor model, and random forest model. Results showed that the AUC values of the five models in the training and test cohorts were 0.905 (95% confidence interval [CI] 0.826-0.984), 0.870 (95% CI 0.742-0.996), 0.910 (95% CI 0.717-0.989), 0.853 (95% CI 0.717-0.989), 0.885 (95% CI 0.787-0.983), 0.628 (95% CI 0.353-0.903), 0.925 (95% CI 0.873-0.977), 0.638 (95% CI 0.416-0.861), 0.980 (95% CI 0.954-1.000), and 0.821 (95% CI 0.673-0.968). In summary, the LR model showed the best performance and had ideal diagnostic performance. Conclusion Enhanced CT-based image omics features can improve the accuracy of the differential diagnosis of ccRCC and RO.
Differential diagnosis of renal eosinophilic adenoma and renal clear cell carcinoma based on enhanced CT image omics model
Objective This study aimed to explore the value of enhanced CT imaging model combined with machine learning in the differential diagnosis of renal clear cell carcinoma (ccRCC) and renal eosinophilic adenoma (RO). Methods Retropective cohort study was conduted. A total of 100 patients with renal cancer diagnosed pathologically admitted to Shanxi Cancer Hospital from June 2013 to July 2022 were included, including 63 males and 37 females, aged 42-81 (60.8±8.7) years. Of the 100 patients, 75 were pathologically confirmed as ccRCC patients, including 52 males and 23 females, aged 42-79 (59.9±8.5) years. There were 25 RO patients in the RO group, including 11 males and 14 females, aged 46-81 (63.5±8.8) years. The patients were classified into test cohort (30%) and training cohort (70%) with a random method, and all of them underwent enhanced CT examination. Then, 1 409 imaging features were extracted from portal venous CT images, and variance threshold, SelectKBest, and least absolute shrinkage and selection operator algorithms were used in feature extraction. We constructed an image omics model and drew the receiver operating characteristic curve. Sensitivity, specificity, accuracy, and F1-score were used in evaluating the performance of different imaging models. Results No statistically significant difference in age was found between the groups (t=-1.82, P=0.072). The gender showed statistically significant difference between the groups (χ2=5.16, P=0.023). The 12 most relevant features were selected from 1 409 image omics features, and five models were established to differentiate between ccRCC and RO: support vector machine, logistic regression (LR) model, decision tree model, K-nearest neighbor model, and random forest model. Results showed that the AUC values of the five models in the training and test cohorts were 0.905 (95% confidence interval [CI] 0.826-0.984), 0.870 (95% CI 0.742-0.996), 0.910 (95% CI 0.717-0.989), 0.853 (95% CI 0.717-0.989), 0.885 (95% CI 0.787-0.983), 0.628 (95% CI 0.353-0.903), 0.925 (95% CI 0.873-0.977), 0.638 (95% CI 0.416-0.861), 0.980 (95% CI 0.954-1.000), and 0.821 (95% CI 0.673-0.968). In summary, the LR model showed the best performance and had ideal diagnostic performance. Conclusion Enhanced CT-based image omics features can improve the accuracy of the differential diagnosis of ccRCC and RO.
万方数据
