基于CT影像组学和临床特征构建肝癌新辅助/转化治疗后序贯手术切除病理学完全缓解（pCR）预测模型，并探讨其预测价值。

回顾性分析2019年12月至2022年12月南方医科大学南方医院收治的74例新辅助/转化治疗后序贯手术切除的肝癌患者临床资料。患者均签署知情同意书，符合医学伦理学规定。其中男65例，女9例；年龄 22~74岁，中位年龄52岁。依据术后肿瘤标本（80枚）按随机数表法分训练集（40枚）与验证集（40枚）两组。基于患者CT影像组学提取、筛选特征指标并得出影像组学评分（Rad-score）。基于训练集临床因素Logistic多因素分析和CT影像组学构建pCR列线图预测模型。采用C-index和Hosmer-Lemeshow拟合优度检验评价列线图预测模型的一致性。采用Bootstrap法进行内部验证评估校准曲线准确度。采用绘制ROC曲线下面积（AUC）评价列线图预测模型的区分度。绘制临床决策曲线分析（DCA）评价预测模型的临床效用。绘制临床影响曲线（CIC）评价预测模型临床有效率。Delong检验预测模型间的诊断效能。绘制Rad-score可视化图形展示影像组学模型的诊断性能。

Logistic多因素分析显示，治疗后AFP阳性是pCR独立影响因素（OR=5.250，95%CI：1.069~25.789；P<0.05）。经Radiomics提取共4 148个影像特征，采用Lasso分析和十折交叉验证筛选出11个影像特征。AFP与Rad-score联合构建pCR列线图预测模型（联合模型）。联合模型的校准曲线在预测的pCR概率与实际发生概率有较好的一致性，训练集和验证集的C-index分别为0.887、0.895，Hosmer-Lemeshow拟合优度较好（χ²=5.96，4.78；P>0.05）。训练集和验证集Bootstrap法内部验证显示，联合模型校准曲线的平均绝对误差分别为0.063、0.040，曲线准确度良好。DCA和CIC分析显示联合模型具有较高的临床效用和较好的临床有效率。训练集和验证集中，联合模型预测的AUC分别为0.887、0.895。训练集中临床模型、影像组学模型的AUC分别为0.887、0.667、0.857，联合模型诊断效能明显优于临床模型（Z=2.797，P=0.005 2）；验证集联合模型诊断效能亦明显优于临床模型（Z=2.027，P=0.042 7）。影像组学Rad-score瀑布图示预测错误的比例不高，可视化良好。

基于CT影像组学和AFP构建的肝癌新辅助/转化治疗后序贯手术切除pCR的联合模型预测能力良好，诊断效能最佳。

To construct a prediction model for pathologic complete response (pCR) in liver cancer patients after neoadjuvant/conversion therapy followed by sequential surgical resection based on CT radiomics and clinical features, and to evaluate its predictive value.

Clinical data of 74 liver cancer patients undergoing neoadjuvant/conversion therapy, followed by sequential surgical resection in Nanfang Hospital of Southern Medical University from December 2019 to December 2022 were retrospectively analyzed. The informed consents of all patients were obtained and the local ethical committee approval was received. Among them, 65 patients were male and 9 female, aged from 22 to 74 years, with a median age of 52 years. According to postoperative tumor specimens (n=80), they were randomly divided into the training (n=40) and validation sets (n=40). Feature indexes were extracted and screened based on CT radiomics of the patients, and the radiomics score (Rad-score) was obtained. A nomogram prediction model for pCR was constructed based on multivariate Logistic analysis of clinical factors and CT radiomics in the training set. The consistency of nomogram prediction model was evaluated by C-index and Hosmer-Lemeshow goodness-of-fit test. Bootstrap method was used for internal verification to evaluate the accuracy of calibration curve. The area under the ROC curve (AUC) was utilized to evaluate the degree of discrimination of nomogram prediction model. Decision curve analysis (DCA) was drawn to evaluate clinical utility of the prediction model. Clinical impact curve (CIC) was delineated to evaluate clinical effectiveness of the prediction model. The diagnostic efficiency of the prediction model was assessed by Delong test. The visual graph of Rad-score was drawn to display the diagnostic performance of the radiomics model.

Multivariate Logistic analysis showed that positive AFP after treatment was an independent influencing factor of pCR (OR=5.250, 95%CI: 1.069-25.789; P<0.05). A total of 4 148 radiomic features were extracted by Radiomics, and 11 radiomic features were screened by Lasso analysis and 10-fold cross-validation. AFP combined with Rad-score were used to construct a nomogram prediction model for pCR (combined model). The calibration curve of the combined model yielded high consistency between the predicted and actual pCR probability. The C-index in the training and validation sets was 0.887 and 0.895. Hosmer-Lemeshow test showed a high goodness of fit (χ²=5.96, 4.78; both P>0.05). Internal verification using Bootstrap method in the training and validation sets demonstrated that the average absolute error of the calibration curve of the combined model was 0.063 and 0.040, with high accuracy. DCA and CIC analyses showed that the combined model possessed high clinical utility and clinical effectiveness rate. In the training and validation sets, the AUC of the combined model was 0.887 and 0.895. The AUC of clinical and radiomics models in the training set was 0.887, 0.667 and 0.857, respectively. The diagnostic efficiency of combined model was significantly better than that of clinical model (Z=2.797, P=0.005 2). The diagnostic efficiency of combined model in the validation set was also significantly better than that of clinical model (Z=2.027, P=0.042 7). Rad-score waterfall plot revealed the proportion of incorrect prediction was not high, and the visualization effect was good.

The combined model based on CT radiomics and AFP possesses high prediction value for pCR after neoadjuvant/conversion therapy followed by sequential surgical resection for liver cancer and yields the optimal diagnostic efficiency.