摘要目的:探讨CT影像组学特征结合临床风险因素建立的组合模型对高级别浆液性卵巢癌（HGSOC）患者初次肿瘤细胞减灭术（PDS）预后的评估价值。方法:回顾性队列研究。纳入山西医科大学附属肿瘤医院2014年10月—2022年5月确诊HGSOC并行PDS治疗的患者170例，年龄26~79（55.5±9.7）岁。采用随机抽样法将170例患者以7∶3的比例随机分为训练组（118例）和验证组（52例）。应用阈值方差法、K最佳选择法，以及最小绝对收缩和选择算法，筛选出与预后相关的CT影像组学特征，构建影像组学模型；通过单因素和多因素logistic回归分析筛选出HGSOC预后相关的临床风险因素，构建临床模型；将筛选出的CT影像组学特征联合临床风险因素结合梯度提升机器学习方法建立组合模型。在训练集和验证集中，通过受试者操作特征曲线下面积（AUC）和决策曲线分析评估3种模型对患者预后的预测效能。结果:训练组与验证组患者的临床资料比较，除PDS后疾病残留状态外，其他临床指标差异均无统计学意义（ P值均>0.05）。2组内的生存与死亡患者比较：年龄、甲胎蛋白水平、癌胚抗原水平、糖类抗原199水平差异均无统计学意义（ P值均>0.05），人附睾蛋白4水平、国际妇产科学联合会（FIGO）分期、PDS后疾病残留状态差异均有统计学意义（ P值均<0.05），糖类抗原125水平、淋巴结转移状态在验证组中差异均有统计学意义（ P值均<0.05）、在训练组中差异均无统计学意义（ P值均>0.05）。用筛选出的6个CT影像组学特征构建影像组学模型，影像组学模型预测HGSOC患者PDS术后预后的AUC在训练组和验证组分别为0.93（95% CI：0.90~0.96）和0.81（95% CI：0.54~0.72）；用筛选出的3个临床风险因素（HE4、PDS后疾病残留状态和FIGO分期）构建临床模型，临床模型在训练组和验证组预测预后的AUC分别为0.98（95% CI：0.97~0.99）和0.89（95% CI：0.86~0.97）；联合CT影像组学特征和临床风险因素构建的组合模型在训练和验证组预测预后的AUC分别为0.99（95% CI：0.98~1.00）和0.95（95% CI：0.83~0.98）。3种模型中，组合模型对训练组和验证组患者的预后预测具有较高的一致性，其决策收益率大于影像组学模型和临床模型。 结论:相比于临床模型和CT影像组学模型，CT影像组学特征与临床风险因素结合建立的组合模型对HGSOC患者PDS后的预后预测效能更好。

abstractsObjective:CT-radiomics features were combined with clinical risk factors to establish a combination model for evaluating the prognosis of primary debulking surgery in patients with high-grade serous ovarian cancer (HGSOC).Methods:We conducted a retrospective cohort study including 170 patients aged 26-79(55.5±9.7) years. They were diagnosed with HGSOC and underwent primary debulking surgery from October 2014 to May 2022. Based on random sampling numbers, all 170 patients were divided into training (118 patients) and validation (52 patients) groups in a 7∶3 ratio. Three feature-screening methods including variance threshold, select KBest, and least absolute shrinkage and selection operator were used to screen out prognostic-related CT-radiomics features and construct radiomics model. A clinical model was built using univariate and multivariate logistic-regression analyses based on the clinical risk factors related to prognosis. A combination model was established by combining the selected radiomics features with clinical risk factors and gradient-boosting machine learning. In the training and validation groups, the model's performance in predicting patient prognosis was evaluated by the receiver-operating characteristic curve area under the curve and decision-curve analysis.Results:Except for residual disease status after PDS, there were no statistically significant differences in clinical data existed between the training and validation groups (all P values > 0.05). No significant differences in age, alpha-fetoprotein level, carcinoembryonic antigen level, and carbohydrate antigen 199 level were observed between the two groups (all P values > 0.05). However, statistically significant differences were found in human epididymal protein 4 level, International Federation of Gynecology and Obstetrics (FIGO) stage, and residual disease status after primary debulking surgery (all P values < 0.05). Statistically significant variations existed in carbohydrate antigen 125 levels and lymph-node metastasis status between patients who survived and those who died within the validation group (all P values < 0.05 ). However, no statistical significance was observed within the training group (all P values > 0.05 ). Three-feature-selection methods were applied to filter out six radiomics features, which were used to construct the model of radiomics. The radiomics model AUC in the training and validation of the model group were 0.93 (95% CI: 0.90-0.96) and 0.81(95% CI: 0.54-0.72), respectively. We filtered three clinical risk factors, including HE4, residual disease status after primary debulking surgery surgery, and FIGO stage build clinical models. The clinical models in the training and validation sets of AUC were 0.98 (95% CI: 0.97-0.99) and 0.89 (95% CI: 0.86-0.97). The AUC of the combination model, which combined radiomics features and clinical risk factors, was 0.99 (95% CI: 0.98-1.00) and 0.95 (95% CI: 0.83-0.98) in the training and validation groups, respectively. Among the three models, the combination model exhibited high-level consistency in prognostic prediction between the training and validation groups, surpassing the radiomics and clinical models in terms of decision-return rate. Conclusion:Compared with clinical and CT radiomics models, the combination model established by combining CT radiomics features and clinical risk factors has a better prediction efficiency for the prognosis of patients with HGSOC after primary debulking surgery.