摘要目的:探讨增强CT熵特征联合传统影像征象的组合模型鉴别诊断胸腺上皮性肿瘤（TET）危险程度的临床价值。方法:回顾性分析2008年10月至2021年5月在江门市中心医院和中山大学附属第五医院经手术和组织病理学检查结果证实的178例TET患者的临床资料[男性83例、女性95例；年龄（52.7±12.4）岁，范围26~83岁]，按照组织病理学亚型分为低危组（A、AB和B1型）和高危组（B2和B3型）。将全部患者进一步分为训练集（ n=86）、内部验证集（ n=51）、外部验证集（ n=41），其中内部验证集和外部验证集合称为全部验证集（ n=92），训练集和验证集分别用于预测模型的过程构建和效能评价。记录TET患者的临床特征，分析病灶的传统CT征象。应用MATLAB R2016平台的开发软件，在增强CT静脉期图像上定量提取、筛选熵特征。通过Mann-Whitney U检验筛选有鉴别价值的熵特征，采用极限学习机（ELM）分类算法计算熵特征权重和熵标签值。采用多因素Logistic回归分析分别构建临床模型、熵模型和组合模型，并采用受试者工作特征曲线对比3个预测模型的诊断效能。 结果:178例TET患者中，低危组83例[男性38例、女性45例；年龄（52.8±12.4）岁，范围26~83岁]；高危组95例[男性45例、女性50例；年龄（52.0±12.0）岁，范围27~80岁]。单因素分析结果显示，训练集CT征象中周围侵犯在2组间的差异有统计学意义（ χ2=5.108， P=0.024）。共提取初始熵特征1 680个，最终筛选到21个核心熵特征，通过ELM计算得出训练集中低危组熵标签值为0.519± 0.21，明显低于高危组（0.997±0.23），2组间差异有统计学意义（ t=-9.747， P<0.001）。熵模型在训练集、内部验证集、外部验证集和全部验证集的曲线下面积（AUC）分别为0.929(95% CI：0.876~0.983）、0.832(95% CI：0.723~0.941）、0.802(95% CI：0.666~0.939）、0.803(95% CI：0.715~0.890）。多因素Logistic回归分析结果显示，周围侵犯（ OR=6.343；95% CI：1.009~36.604； P=0.039）和熵标签值（ OR=20.145；95% CI：5.887~68.936； P<0.001）是预测TET危险程度的独立危险因素，二者共同构建的组合模型在训练集、内部验证集、外部验证集和全部验证集的AUC分别为0.941(95% CI：0.894~0.987）、0.871(95% CI：0.775~0.968）、0.819(95% CI：0.689~0.949）、0.840(95% CI：0.761~0.919）。 结论:基于胸部增强CT图像的熵特征可以定量评估TET的危险程度；周围侵犯和熵标签值构建的组合模型的诊断效能最高，可以精准指导TET患者的术前治疗策略。

abstractsObjective:To investigate the clinical value of a combined model of enhanced CT-based entropy features and traditional imaging features for the differential diagnosis of risk status in patients with thymic epithelial tumor (TET).Methods:A retrospective analysis was conducted on the clinical data of 178 TET patients (83 males and 95 females; aged (52.7±12.4) years, ranged from 26 years to 83 years) confirmed by surgical and histopathological results at Jiangmen Central Hospital and the Fifth Affiliated Hospital of Sun Yat-sen University from October 2008 to May 2021 were retrospectively analyzed. They were divided into low-risk (A, AB, and B1 types) and high-risk (B2 and B3 types) groups according to histopathology subtypes. All patients were further divided into a training set ( n=86), an internal validation set ( n=51), and an external validation set ( n=41). The internal and external validation sets were called the total validation set ( n=92). The training and validation sets were used for the process construction and performance evaluation of the predictive model, respectively. The clinical characteristics of TET patients were recorded, and the traditional CT features of lesions were analyzed. Entropy features were extracted and selected from enhanced CT venous phase images by using software developed based on MATLAB R2016 platform. Mann-Whitney U test was used to select valuable entropy features. Extreme learning machine classification algorithm was adopted to calculate the different weights of entropy features and entropy signature value. Clinical model, entropy model, and combined model were constructed through Logistic analysis, and receiver operating characteristic curve was used to compare the diagnostic efficacy of the three predictive models. Results:A total of 83 cases were included in the low-risk group (38 males and 45 females; aged (52.8±12.4) years, ranged from 26 years to 83 years), whereas 95 cases were in the high-risk group (45 males and 50 females; aged (52.0±12.0) years, ranged from 27 years to 80 years). Univariate analysis showed that the difference in peripheral invasion between the two groups in the training set was statistically significant ( χ2=5.108, P=0.024). A total of 1 680 initial entropy features were extracted, and 21 core entropy features were ultimately selected. The entropy signature value of the low-risk group in the training set was 0.519±0.21, which was significantly lower than that of the high-risk group (0.997±0.23). The difference between the two groups was statistically significant ( t=-9.747, P<0.001). The area under curve (AUC) of the entropy model in the training set, internal validation set, external validation set, and total validation set were 0.929(95% CI: 0.876-0.983), 0.832(95% CI: 0.723-0.941), 0.802(95% CI: 0.666-0.939), and 0.803(95% CI: 0.715-0.890), respectively. Results of multivariate Logistic regression analysis showed that peripheral invasion ( OR=6.343; 95% CI: 1.009-36.604; P=0.039) and entropy signature value ( OR=20.145; 95% CI: 5.887-68.936; P<0.001) were independent risk factors for predicting TET risk status. The combined model constructed by the two together had AUCs of 0.941(95% CI: 0.894-0.987), 0.871(95% CI: 0.775-0.968), 0.819(95% CI: 0.689-0.949), and 0.840(95% CI: 0.761-0.919) in the training set, internal validation set, external validation set, and total validation set, respectively. Conclusions:The entropy feature based on enhanced chest-CT images can be used to assess the risk status of TET quantitatively. The combined model constructed by peripheral invasion and entropy signature value showed the highest diagnostic efficacy, which can accurately guide the preoperative treatment strategy of TET patients.