第一部分 基于磁共振影像组学模型预测肝内胆管癌瘤内三级淋巴结构及术后复发研究

摘要

【目的】

探讨与肝内胆管癌（intrahepatic cholangiocarcinoma，iCCA）瘤内三级淋巴结构（tertiary lymphoid structures，TLSs）相关的临床特征、磁共振成像（magnetic resonance imaging，MRI）特征、影像组学特征，构建MRI影像学模型并验证模型对iCCA术后复发的预测价值。

【材料及方法】

回顾性收集2011年6月至2022年7月在中国医学科学院肿瘤医院术后病理证实为iCCA的患者151例，2019年7月至2021年12月在河南省肿瘤医院术后病理证实为iCCA的患者41例。使用ITK-SNAP软件在横轴位T2WI脂肪抑制序列（fat-suppressed T2-weighted imaging，T2WI/FS）、扩散加权成像（diffusion-weighted imaging，DWI）、增强扫描门静脉期（portal venous phase，PVP）图像上手动全层勾画感兴趣区（region of interest，ROI）。使用最大相关最小冗余（maximum relevance and minimum redundancy，mRMR）和最小绝对收缩和选择算子（least absolute shrinkage and selection operator，LASSO）两种方法选择与TLSs相关的特征，并构建影像组学模型。采用单因素和多因素Logistic回归分析与TLSs相关的临床、影像学特征，确定与TLSs相关的独立预测因素。使用曲线下面积（area under curve，AUC）比较不同模型的预测效能，绘制影像组学模型的校准曲线，并使用Hosmer-Lemeshow检验以评估预测危险分层和实际危险分层之间的一致性。采用决策曲线分析（decision curve analysis，DCA）评价影像组学模型的临床应用价值。最后使用Kaplan -Meier方法与Log-rank检验进行TLSs状态和影像组学模型的无复发生存期（recurrence-free survival，RFS）生存分析并绘制生存曲线。

【结果】

共保留287个具有较好观察者间一致性的影像组学特征，最终共筛选出11个最优的影像组学特征，通过系数加权来得出影像组学评分（radiomics score，Rad-score）。单因素和多因素Logistic回归分析结果显示，只有动脉期弥漫高强化是TLSs状态的独立预测因子（OR [95% CI]：3.082[1.987-4.178]）。Rad-score的AUC在训练队列为0.85（95% CI，0.77-0.92）), 内部验证队列V1为0.81（95% CI，0.67-0.94），外部验证队列V2为0.84（95% CI，0.71-0.96）。Hosmer-Lemeshow检验的P值在训练队列、验证队列V1和V2中分别为0.74、0.06和0.74。动脉期弥漫高强化预测TLSs状态的AUC在训练队列、内部验证队列V1、外部验证队列V2中分别为0.59（95% CI，0.50-0.67）、0.52（95% CI，0.43-0.61）、0.66（95% CI，0.52-0.80）。在中国医学科学院肿瘤医院的151例iCCA患者中，肿瘤内TLSs阳性患者的RFS明显优于TLSs阴性患者（中位RFS：35.5；95% CI，12.8-58.2个月vs.中位RFS：9.6；95% CI，7.6-11.6个月）。在训练队列中，低危组（Rad-score≥-0.21）的RFS显著优于高危组（中位RFS：38.7；95% CI：5.4-71.9vs.中位RFS：14.1；95% CI：7.0-21.2）。这在验证队列V1和V2中也得到了证实。

【结论】

1、动脉期弥漫高强化是TLSs状态的独立预测因子。

2、MRI影像组学模型可术前预测iCCA患者的瘤内TLSs状态，并与RFS显著相关。

第二部分 基于CT的影像组学模型预测肝内胆管癌瘤内三级淋巴结构及术后复发研究

摘要

【目的】

探讨与肝内胆管癌（intrahepatic cholangiocarcinoma，iCCA）三级淋巴结构（tertiary lymphoid structures，TLSs）相关的临床特征、影像学特征、CT影像组学特征，构建临床影像学模型、影像组学模型、联合模型，评估三个模型效能并验证模型对iCCA术后复发的预测价值。

【材料及方法】

回顾性纳入2010年11月至2020年8月在中国医学科学院肿瘤医院、2015年5月至2019年11月在河南省肿瘤医院术后病理证实为iCCA的患者86例、30例。使用ITK-SNAP软件在增强CT轴位门静脉期（portal venous phase，PVP）序列上手动全层勾画感兴趣区（region of interest，ROI）。使用最大相关最小冗余（maximum relevance and minimum redundancy，mRMR）和最小绝对收缩和选择算子（least absolute shrinkage and selection operator，LASSO）两种方法选择与TLSs相关的特征，并构建影像组学模型，计算影像组学评分（radiomics score，Rad-score）。采用单因素和多因素Logistic回归分析与TLSs相关的临床特征、影像学特征，建立临床影像学模型，并将影像组学模型及临床影像学模型联合构建联合模型，计算列线图评分。使用受试者工作特征曲线（receiver operating characteristic curve，ROC）的最大约登指数来确定影像组学模型和联合模型的二分类截断值。使用曲线下面积（area under curve，AUC）比较不同模型的预测效能，绘制联合模型的校准曲线，并使用Hosmer-Lemeshow检验以评估预测危险分层和实际危险分层之间的一致性。采用决策曲线分析（decision curve analysis，DCA）评价三个模型的临床应用价值。最后使用Kaplan -Meier方法与log-rank检验进行TLSs状态、影像组学模型、联合模型的生存分析并绘制生存曲线。使用一致性指数（concordance index，C-index）评估 TLSs状态、影像组学模型、联合模型对无复发生存期（recurrence-free survival，RFS）的预测效能。

【结果】

每个患者共提取了107个影像组学特征，经过筛选剔除后，将6个特征通过系数加权建立了用于预测TLSs的影像组学模型。单因素Logistic回归分析结果显示，在训练队列中，动脉期弥漫高强化、动脉期边缘环形高强化、美国癌症联合委员会（American Joint Committee on Cancer，AJCC）第八版TNM分期以及肿瘤直径在TLSs阳性组和阴性组之间存在显著差异（P值分别为<0.001、0.003、0.014、0.039）。多因素Logistic回归分析结果显示共纳入动脉期弥漫高强化和AJCC第八版TNM分期两个因素进入最终的临床影像学模型。将影像组学模型及临床影像学模型联合构建联合模型。在训练队列中，联合模型的效能优于单独的影像组学模型和临床影像学模型（AUC分别为0.85、0.82和0.75），并且在外部验证队列中也得到了验证（AUC分别为0.88、0.86和0.71）。肿瘤内TLSs阳性的患者的RFS明显优于TLSs阴性的患者（中位RFS：46.6个月；95% CI：25.6-67.7个月vs.中位RFS：9.6个月；95% CI：7.9-11.3个月，C-index = 0.598[95% CI：0.537-0.659]，P = 0.014）。在训练队列中，Rad-score≥-0.76（低风险）组的患者RFS显著优于Rad-score＜-0.76（高风险）组的患者（P<0.001，C-index = 0.678）；列线图评分≥-1.16（低风险）组的患者RFS显著优于评分＜-1.16（高风险）组的患者（P<0.001，C-index = 0.723）。且在外部验证队列中得到了验证。

【结论】

1、CT影像组学列线图可作为肿瘤内TLSs状态的术前预测模型，其效能优于单独的影像组学模型或临床影像学模型。

2、术前CT影像组学列线图能够对iCCA患者的RFS进行更准确分层，效能优于术后病理TLSs评估结果。

第三部分 利用免疫血管特征预测肝内胆管癌术后的极早期复发和预后摘要

【目的】

探讨与iCCA极早期复发（very early recurrence，VER）相关的临床特征、影像学特征、病理学特征，构建预测模型，并进一步进行预后分层。

【材料及方法】

回顾性纳入2012年5月至2022年7月期间在中国医学科学院肿瘤医院、2019年7月至2021年12月在河南省肿瘤医院的术后病理证实为iCCA的患者160例、51例，分别作为训练队列和外部验证队列。对临床、影像学及病理学相关变量进行评估和收集。采用单因素和多因素Logistic回归分析来筛选iCCA术后VER的独立预测因素。将筛选出的因素相结合进行分型，采用Kaplan-Meier法和对数秩检验对总生存期（overall survival，OS）进行生存分析。

【结果】

在训练队列中，39例（24.4%）患者出现VER，而121例（75.6%）未出现（非VER组）。在训练队列中，中位OS为40.5个月（95% CI：33.2-47.7个月）。VER组的OS明显短于非VER组（中位OS：14.8个月，95% CI：11.6-18.0个月vs.53.4个月，34.3-72.6个月；P<0.001），且在外部验证队列中也得到了证实（中位OS：22.1个月，95% CI：8.8-35.4个月vs.40.1个月，21.2-59.0个月；P = 0.003）。根据单因素Logistic回归分析，VER组和非VER组之间有四个变量存在显著差异（瘤内三级淋巴结构[tertiary lymphoid structures, TLSs]，P = 0.028；肿瘤分化程度，P = 0.023；微血管侵犯[microvascular invasion, MVI]，P = 0.012；肿瘤直径，P = 0.028）。根据多因素Logistic回归分析，MVI和瘤内TLSs状态是VER的独立预测因素。MVI阳性与VER的发生呈正相关（OR = 2.50；95% CI：1.16-5.18；P = 0.018），而瘤内TLSs阳性与VER的发生呈负相关（OR = 0.43；95% CI：0.19-0.97；P = 0.041）。基于TLSs和MVI状态，将iCCA患者分为四组：TLSs阳性且MVI阴性（TP/MN）；TLSs阴性且MVI阴性（TN/MN）；TLSs阳性且MVI阳性（TP/MP）；TLSs阴性且MVI阳性（TN/MP）。在训练队列中，这四个分型与OS分层显著相关（P<0.001），并且在外部验证队列中也得到了验证（P<0.001）。

【结论】

瘤内TLSs和MVI状态是可切除iCCA患者术后VER的独立预测因素，基于此构建的免疫血管分型与OS显著相关。

Part 1

Prediction of tertiary lymphoid structures and postoperative recurrence of intrahepatic cholangiocarcinoma based on MRI radiomics model

Abstract

[Objective]

To investigate the clinical features, magnetic resonance imaging (MRI) features, and MRI radiomics features related to intra-tumoral tertiary lymphoid structures (TLSs) of intrahepatic cholangiocarcinoma (iCCA), and to construct a radiologic model and verify the predictive value of the model for postoperative recurrence of iCCA.

[Materials and Methods]

A total of 151 patients with iCCA confirmed by postoperative pathology in Cancer Hospital of Chinese Academy of Medical Sciences from June 2011 to July 2022 and 41 patients with iCCA confirmed by postoperative pathology in Henan Cancer Hospital from July 2019 to December 2021 were retrospectively collected. ITK-SNAP software was used to manually contouring the region of interest (ROI) on axial fat-suppressed T2-weighted imaging (T2WI/FS), diffusion-weighted imaging (DWI), and enhanced portal venous phase (PVP) images. Using maximum relevance and minimum redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO) method to select the features related to TLSs and construct the radiomics model. Univariate and multivariate logistic regression analysis were used to analyze the clinical and imaging features associated with TLSs, and to determine the independent predictors of TLSs. The calibration curve of radiomics model was drawn to assess consistency between prediction and observation, accompanied by the Hosmer-Lemeshow test. Decision curve analysis (DCA) was used to evaluate the clinical application value of the radiomics model. Kaplan-Meier method and log-rank test were used to analyze the recurrence-free survival (RFS) of TLSs status and radiomics model, and the survival curves were drawn.

[Results]

A total of 287 radiomics features with good inter-observer agreement were retained, and 11 optimal radiomics features were selected. The radiomics score (Rad-score) was calculated by coefficient weighting. Univariate and multivariate logistic regression analysis showed that only arterial phase diffuse hyperenhancement was an independent predictor of TLSs status (OR [95% CI]: 3.082[1.987-4.178]). The AUCs of the Rad-score in the training cohort, internal validation cohort V1 and external validation cohort V2 were 0.85 (95% CI, 0.77 to 0.92), 0.81(95% CI, 0.67 to 0.94), and 0.84 (95% CI, 0.71 to 0.96), respectively. The P values of Hosmer-Lemeshow test were 0.74, 0.06 and 0.74 in the training cohort, validation cohort V1 and V2, respectively. The AUCs of arterial phase diffuse hyperenhancement for predicting TLSs status were 0.59 (95% CI, 0.50-0.67) in the training cohort, 0.52 (95% CI, 0.43-0.61) in the validation cohort V1, and 0.66 (95% CI, 0.52-0.80) in the validation cohort V2. In the cohort of 151 iCCA patients from Cancer Hospital, Chinese Academy of Medical Sciences, patients with TLSs positive had significantly better RFS than those with TLSs negative (median RFS: 35.5; 95% CI, 12.8-58.2 months vs. 9.6; 95% CI, 7.6-11.6 months). In the training cohort, the RFS of the low-risk group (Rad-score≥-0.21) was significantly better than that of the high-risk group (median RFS: 38.7; 95% CI: 5.4-71.9 vs. median RFS: 14.1; 95% CI: 7.0-21.2). This was also confirmed in validation cohorts V1 and V2.

[Conclusions]

1. Arterial phase diffuse hyperenhancement is an independent predictor of TLSs status.

2. The MRI radiomics model can predict the status of intra-tumoral TLSs in patients with iCCA preoperatively, and it is significantly correlated with RFS.

Part 2

Prediction of tertiary lymphoid structures and postoperative recurrence of intrahepatic cholangiocarcinoma using a CT-based radiomics model

Abstract

[Objective]

To explore the clinical features, imaging features, and CT radiomics features related to tertiary lymphoid structures (TLSs) in intrahepatic cholangiocarcinoma (iCCA), construct a clinical imaging model, a radiomics model, and a combined model, evaluate the efficacy of the three models, and verify the predictive value of the models for the postoperative recurrence of iCCA.

[Materials and Methods]

A total of 86 patients with iCCA confirmed by postoperative pathology in the Cancer Hospital, Chinese Academy of Medical Sciences from November 2010 to August 2020 and 30 patients in Henan Cancer Hospital from May 2015 to November 2019 were retrospectively included. The ITK-SNAP software was used to manually delineate the region of interest (ROI) layer by layer on the axial portal venous phase (PVP) sequence of enhanced CT. Two methods, namely maximum relevance and minimum redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO), were used to select the features related to TLSs, construct a radiomics model, and calculate the radiomics score (Rad-score). Univariate and multivariate logistic regression analyses were performed on the clinical features and imaging features related to TLSs to establish a clinical imaging model. The radiomics model and the clinical imaging model were combined to construct a combined model, and the nomogram score was calculated. The maximum Youden index of the receiver operating characteristic curve (ROC) was used to determine the binary classification cut-off value of the radiomics model and the combined model. The area under the curve (AUC) was used to compare the predictive efficacy of different models. The calibration curve of the combined model was plotted, and the Hosmer-Lemeshow test was used to evaluate the consistency between the prediction and observation. Decision curve analysis (DCA) was used to evaluate the clinical application value of the three models. Finally, the Kaplan-Meier method and log-rank test were used for the survival analysis of the TLSs status, radiomics model, and combined model, and the survival curves were plotted. The concordance index (C-index) was used to evaluate the predictive efficacy of the TLSs status, radiomics model, and combined model for recurrence-free survival (RFS).

[Results]

A total of 107 radiomics features were extracted from each patient. After screening and elimination, six features were weighted by coefficients to establish a radiomics model for predicting TLSs. The results of univariate logistic regression analysis showed that in the training cohort, there were significant differences in arterial diffuse hyperenhancement, arterial peripheral rim enhancement, American Joint Committee on Cancer (AJCC) 8th TNM staging, and tumor diameter between the TLSs positive group and the negative group (P values were < 0.001, 0.003, 0.014, and 0.039 respectively). The results of multivariate logistic regression analysis showed that two factors, namely arterial diffuse hyperenhancement and the AJCC 8th TNM staging, were included in the final clinical imaging model. The radiomics model and the clinical imaging model were combined to construct a combined model. In the training cohort, the efficacy of the combined model was better than that of the independent radiomics model and the clinical imaging model (AUCs were 0.85, 0.82, and 0.75 respectively), and it was also verified in the external validation cohort (AUCs were 0.88, 0.86, and 0.71 respectively). The RFS of patients with intra-tumoral positive TLSs was significantly better than that of patients with negative TLSs (median RFS: 46.6 months; 95% CI: 25.6-67.7 months vs. median RFS: 9.6 months; 95% CI: 7.9-11.3 months, C-index = 0.598[95% CI: 0.537-0.659], P = 0.014). In the training cohort, the RFS of patients in the group with Rad-score≥-0.76 (low risk) was significantly better than that of patients in the group with Rad-score<-0.76 (high risk) (P<0.001, C-index = 0.678); the RFS of patients in the group with a nomogram score≥- 1.16 (low risk) was significantly better than that of patients in the group with a score<- 1.16 (high risk) (P<0.001, C-index = 0.723). And it was verified in the external validation cohort.

[Conclusions]

1. The CT radiomics nomogram can serve as a preoperative predictive model for the intra-tumoral TLSs status, and its efficacy is better than that of the independent radiomics model or the clinical imaging model.

2. The preoperative CT radiomics nomogram can stratify the RFS of iCCA patients more accurately, and its efficacy is better than the postoperative pathological evaluation results of TLSs.

Part 3

Using immunovascular characteristics to predict very early recurrence and prognosis of resectable intrahepatic cholangiocarcinoma

Abstract

[Objective]

To explore the clinical, imaging and pathological characteristics associated with very early recurrence (VER) of intrahepatic cholangiocarcinoma, construct a prediction model, and further conduct prognostic stratification.

[Materials and Methods]

A total of 160 patients and 51 patients with intrahepatic cholangiocarcinoma confirmed by postoperative pathology in the Cancer Hospital, Chinese Academy of Medical Sciences from May 2012 to July 2022 and in Henan Cancer Hospital from July 2019 to December 2021 were retrospectively included as the training cohort and the external validation cohort respectively. The clinical, imaging and pathological related variables were evaluated and collected. Univariate and multivariate logistic regression analyses were used to screen the independent predictive factors of VER in intrahepatic cholangiocarcinoma. The selected factors were combined for classification, and the Kaplan-Meier method and log-rank test were used for survival analysis of overall survival (OS).

[Results]

In the training cohort, 39 patients (24.4%) had VER, while 121 patients (75.6%) did not (non-VER group). In the training cohort, the median OS was 40.5 months (95% CI: 33.2-47.7 months). The OS of the VER group was significantly worse than that of the non-VER group (median OS: 14.8 months, 95% CI: 11.6-18.0 months vs. 53.4 months, 34.3-72.6 months; P <0.001), and this was also confirmed in the external validation cohort (median OS: 22.1 months, 95% CI: 8.8-35.4 months vs. 40.1 months, 21.2-59.0 months; P = 0.003). According to the univariate logistic regression analysis, there were four variables significantly different between the VER group and the non-VER group (intra-tumoral tertiary lymphoid structures [TLSs], P = 0.028; tumor differentiation degree, P = 0.023; microvascular invasion [MVI], P = 0.012; tumor diameter, P = 0.028). According to the multivariate logistic regression analysis, the MVI and intra-tumoral TLSs status were independent predictive factors of VER. Positive MVI was positively correlated with the occurrence of VER (OR = 2.5; 95% CI: 1.16-5.18; P = 0.018), while positive intra-tumoral TLSs was negatively correlated with the occurrence of VER (OR = 0.43; 95% CI: 0.19-0.97; P = 0.041). Based on the TLSs and MVI status, patients with intrahepatic cholangiocarcinoma were divided into four groups: TLSs positive and MVI negative (TP/MN); TLSs negative and MVI negative (TN/MN); TLSs positive and MVI positive (TP/MP); TLSs negative and MVI positive (TN/MP). In the training cohort, these four classifications were significantly correlated with OS stratification (P <0.001), and this was also verified in the external validation cohort (P <0.001).

[Conclusions]

The intra-tumoral TLSs and MVI status are independent predictive factors for postoperative VER in patients with resectable iCCA, and the immunovascular classification constructed based on them is significantly correlated with OS.