第一部分 基于T2WI影像组学的逻辑回归在直肠癌侧方淋巴结转移预测中的应用研究

研究背景：直肠癌患者术后局部复发的重要原因之一是侧方淋巴结转移 (lateral pelvic lymph node metastasis, LPLNM)，而新辅助放化疗联合侧方淋巴结清扫术 (lateral pelvic lymph node dissection, LPLD) 被证实可有效降低直肠癌术后复发率。目前，公认的LPLNM临床诊断依据为磁共振 (magnetic resonance imaging, MRI) 中侧方淋巴结 (lateral pelvic lymph node, LPLN) 的最大短轴直径 (short-axis diameter, SAD) 和恶性特征。然而，此诊断方法存在较高的病理阴性清扫率或较强的主观性。因此，有必要开发性能更优的LPLNM预测模型，以制定个体化的LPLNM治疗方案、避免不必要的治疗并发症。

研究目的：本研究基于T2WI (T2-weighted imaging) 影像组学的逻辑回归 (logistic regression, LR) 构建直肠癌LPLNM的预测模型，并在独立测试队列中评估模型的预测性能，旨在评价MRI影像组学方法在预测直肠癌LPLNM方面的应用价值，并为第二部分研究奠定基础。

研究方法：本研究回顾性收集了263例在本单位行 (total mesorectal excision, TME) 联合选择性LPLD的直肠癌患者的基线临床病理信息及盆腔MRI图像资料。经过筛选，共有148例病例纳入分析，按照7：3的比例随机划分为训练队列及独立测试队列。首先，采用LR筛选与LPLNM相关的临床病理特征，并构建临床预测模型。随后，在T2WI图像中提取直肠癌原发灶（下称 "T"）及LPLN的影像组学特征，并采用最小冗余最大相关算法进行特征降维。采用LR分别构建基于T、LPLN及T联合LPLN（下称 "T-LPLN"）影像组学特征的预测模型；选取具有最高受试者工作特征曲线下面积 (area under the receiver operating characteristic curve, AUC) 的模型作为最佳影像组学预测模型。最后，基于临床病理特征及最佳影像组学预测模型，采用LR构建临床-影像组学联合预测模型。比较临床预测模型、最佳影像组学模型及临床-影像组学联合预测模型的AUC值，并评估各模型在最佳分类阈值下的预测表现，据此筛选出最优预测模型。最优预测模型的校准度及临床有效性采用校准曲线及决策曲线分析 (decision curve analysis, DCA) 进行评价。

研究结果：在临床病理特征中，基线癌胚抗原水平、环周切缘状态及LPLN的最大 SAD与LPLNM显著相关，基于此三种特征构建临床预测模型。基于T-LPLN的影像组学预测模型与单独基于T或LPLN影像组学特征的模型相比，在训练及测试队列中的AUC值均最高（分别为0.78及0.73），故将其作为最佳影像组学预测模型。与临床预测模型及最佳影像组学预测模型相比，临床-影像组学联合预测模型在训练及测试队列中均具有最高的ACU值（分别为0.89及0.84）；此外，在最佳分类阈值下，临床-影像组学联合预测模型相较于其它模型，训练及测试队列中的阳性预测值增加约10-20%，同时假阳性率及假阴性率降低，因此被确定为最优预测模型。校准曲线及DCA显示临床-影像组学联合预测模型具有较好的校准度及临床有效性。

研究结论：1、基于T2WI 影像组学的LR，联合临床特征及T、LPLN的影像组学特征所建立的LPLNM预测模型具有较好的区分度、校准度及临床有效性；2、MRI影像组学方法在预测直肠癌LPLNM方面具有良好的应用潜力。

第二部分 基于多模态MRI影像组学的集成学习在直肠癌侧方淋巴结转移预测中的应用研究

研究背景：尽管T2WI (T2-weighted imaging) 影像组学方法在预测直肠癌侧方淋巴结转移 (lateral pelvic lymph node metastasis, LPLNM) 方面具有较好的性能，但单模态图像中可利用的疾病信息有限，而多模态图像可提供不同成像模态间的互补信息，进而更为全面和深入地反映疾病特征。此外，随着图像模态种类的增加，影像组学特征的数量以及特征间的非线性关系随之增多。在此情况下，基于单一算法的预测模型无法充分识别和学习数据中的潜在联系，而集成学习 (ensemble learning, EL) 作为一种先进的机器学习 (machine learning, ML) 策略，可通过整合多个单一模型的预测结果，更有效地揭示数据的复杂性和异质性。因此，基于多模态磁共振 (magnetic resonance imaging, MRI) 影像组学的EL有望进一步提高LPLNM预测模型的性能。

研究目的：本研究在第一部分研究的基础上，采用EL构建基于多模态MRI影像组学的LPLNM预测模型，并在多中心、前瞻性队列中进行模型测试，旨在探索基于多模态MRI影像组学的EL在直肠癌LPLNM预测中的应用潜力。

研究方法：本研究纳入了397例行TME联合选择性LPLD的直肠癌患者，包括本单位的回顾性队列262例（“队列1”）、前瞻性队列89例（“队列2”）及外单位的回顾性队列46例（“队列3”）。将队列1作为训练队列，队列2及队列3合并作为独立测试队列。共有210例病例纳入分析，其中训练队列147例，独立测试队列63例。在T2WI及DWI (diffusion weighted imaging) 中提取直肠癌原发灶（下称 "T"）及LPLN的影像组学特征，采用孤立森林及距离相关性算法进行特征预处理及降维。使用包括类别提升 (categorical boosting, CatBoost) 在内的六种ML算法分别构建分类器，综合考虑各分类器在测试队列中的受试者工作特征曲线下面积 (area under the receiver operating characteristic curve, AUC) 及各自算法优势，筛选出用于建模的最优分类器。构建基于T、LPLN及其组合的多模态影像组学预测模型，比较各模型的AUC值及在最佳分类阈值下的预测表现，据此筛选出最佳影像组学预测模型，并在此基础上结合临床病理特征构建临床-影像组学联合预测模型。比较最佳影像组学预测模型及临床-影像组学联合预测模型的AUC值，及二者在最佳分类阈值下的预测表现，据此筛选出最优预测模型。最优预测模型的校准度及临床有效性采用校准曲线及决策曲线分析 (decision curve analysis, DCA) 评价。

研究结果：选取CatBoost作为最优分类器，构建多模态影像组学预测模型。与基于单独T或LPLN的影像组学预测模型相比，基于二者组合的模型在测试队列中具有更高的AUC、特异度 (Specificity, SPE) 及阳性预测值 (positive predictive value, PPV)，故将其作为最佳影像组学预测模型。在测试队列中，临床-影像组学联合预测模型与最佳影像组学预测模型相比具有更高的AUC值 (0.96比0.92，P < 0.05)，且敏感度、SPE、PPV及阴性预测值均有提高，故将临床-影像组学联合预测模型选作LPLNM的最优预测模型。校准曲线及DCA显示临床-影像组学联合预测模型具有良好的校准度及临床有效性。

研究结论：1.基于多模态MRI影像组学的EL在预测直肠癌LPLNM的临床实践中具有可行性；采用此方法构建的联合临床特征及T、LPLN的多模态影像组学特征的LPLNM预测模型具有良好的区分度、校准度及临床有效性，并在多中心及前瞻性测试队列中表现出良好的泛化能力。2.基于多模态MRI影像组学的EL在精准预测直肠癌LPLNM方面的具有良好的应用潜力；3.CatBoost在处理复杂、高维度的多模态MRI影像组学数据方面表现出色，能够有效处理数据中的非线性关系，有助于提高模型的泛化能力和准确性。

Part 1: Application of T2WI Radiomics-based Logistic Regression in Predicting Lateral Pelvic Lymph Node Metastasis in Rectal Cancer

Background: Lateral pelvic lymph node metastasis (LPLNM) is a significant cause of local recurrence after surgery in rectal cancer patients, and neoadjuvant chemoradiotherapy combined with lateral pelvic lymph node dissection (LPLD) can effectively reduce postoperative recurrence rates. Currently, the recognized clinical diagnostic criteria for LPLNM are the maximum short-axis diameter (SAD) and malignant features of lateral pelvic lymph nodes (LPLN) in magnetic resonance imaging (MRI). However, this diagnostic method has a high rate of pathologically negative result or subjectivity. Therefore, it is necessary to develop a more effective predictive model for LPLNM to tailor individualized treatment plans and avoid unnecessary treatment complications.

Objective: This study aims to construct a predictive model for LPLNM based on T2-weighted imaging (T2WI) radiomics and logistic regression (LR), evaluate the predictive performance of the model in an independent test cohort, assess the application value of MRI radiomics in predicting LPLNM in rectal cancer, and lay the foundation for the second part of the study.

Methods: This retrospective study collected baseline clinicopathological information and pelvic MRI data of 263 rectal cancer patients who underwent total mesorectal excision (TME) combined with selective LPLD at our institution. After screening, a total of 148 cases were analyzed, with participants randomly allocated into training and independent test groups at a ratio of 7:3. Initially, LR was used to select clinicopathological features related to LPLNM and construct a clinical prediction model. Subsequently, radiomic features of the primary rectal cancer lesion (referred to as "T") and LPLN were extracted from T2WI images, and feature dimensionality reduction was performed using the minimum redundancy maximum relevance algorithm. Radiomic features from T, LPLN, and T combined with LPLN (referred to as "T-LPLN") were incorporated into LR to construct radiomics prediction models. The model with the highest area under the receiver operating characteristic curve (AUC) was selected as the optimal radiomics prediction model. Finally, a clinical-radiomics combined prediction model was constructed using LR based on clinicopathological features and the optimal radiomics prediction model. Comparisons were made between the AUC values of the clinical prediction model, the optimal radiomics model, and the combined clinical-radiomics prediction model. Additionally, the performance of each model was evaluated at the optimal classification threshold. The optimal prediction model was selected based on the AUC values and predictive performance at the optimal classification threshold of each model. Calibration curves and decision curve analysis (DCA) were used to evaluate the calibration and clinical effectiveness of the optimal prediction model.

Results: In clinicopathological features, the baseline levels of carcinoembryonic antigen, circumferential resection margin status, and the maximum SAD of LPLN were significantly associated with LPLNM, and a clinical prediction model was constructed based on these three features. The radiomics prediction model based on T-LPLN had the highest AUC values in the training and test cohorts compared to models based solely on T or LPLN radiomic features (0.78 and 0.73, respectively), making it the optimal radiomics prediction model. Compared to the clinical prediction model and the optimal radiomics model, the clinical-radiomics combined prediction model had the highest AUC values in the training and test cohorts (0.89 and 0.84, respectively). Additionally, at the best classification threshold, the clinical-radiomics combined prediction model increased the positive predictive value by approximately 10-20% in both the training and test cohorts, while reducing false positive and false negative rates, thus being identified as the optimal prediction model. Calibration curves and DCA showed that the clinical-radiomics combined prediction model had good calibration and clinical effectiveness.

Conclusion: 1. The prediction model established by combining clinical features with radiomic features of T and LPLN using LR based on T2WI radiomics exhibits good discriminative ability, calibration, and clinical effectiveness in predicting LPLNM. 2. MRI radiomics methods show great potential in predicting LPLNM in rectal cancer.

Part 2: Application of Multimodal MRI Radiomics-Based Ensemble Learning in Predicting Lateral Pelvic Lymph Node Metastasis in Rectal Cancer

Background: While T2-weighted imaging (T2WI) radiomics techniques have demonstrated strong predictive capabilities for detecting lateral pelvic lymph node metastasis (LPLNM) in rectal cancer, the amount of disease information available in single-modal images is constrained. Multimodal images can provide complementary information between different imaging modalities, thus more comprehensively and deeply reflecting the disease characteristics. Additionally, with the increase in image modalities, the number of radiomic features and the nonlinear relationships between features also increase. In this context, prediction models based on a single algorithm cannot fully identify and learn the potential connections in the data. Ensemble learning (EL), as an advanced machine learning (ML) strategy, integrates the predictions of multiple individual models to more effectively reveal the complexity and heterogeneity of the data. Therefore, EL based on multimodal magnetic resonance imaging (MRI) radiomics is expected to further improve the performance of LPLNM prediction models.

Objective: Building on the first part of the study, this research aims to use EL to construct an LPLNM prediction model based on multimodal MRI radiomics and test it in a multicenter, prospective cohort. The goal is to explore the application potential of EL based on multimodal MRI radiomics in predicting LPLNM in rectal cancer.

Methods: This study included 397 rectal cancer patients who underwent total mesorectal excision combined with selective lateral pelvic lymph node dissection, comprising a retrospective cohort of 262 cases from our institution ("Cohort 1"), a prospective cohort of 89 cases from our institution ("Cohort 2"), and a retrospective cohort of 46 cases from an external institution ("Cohort 3"). Cohort 1 was used as the training cohort, and Cohorts 2 and 3 were merged as an independent test cohort. A total of 210 cases were included in the analysis, with 147 cases in the training cohort and 63 cases in the independent test cohort. Radiomic features of the primary rectal cancer lesion (referred to as "T") and LPLN were extracted from T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI). Feature preprocessing and dimensionality reduction were performed using isolation forest and distance correlation algorithms. Six ML algorithms, including categorical boosting (CatBoost), were used to build classifiers, and the optimal classifier for modeling was selected based on the area under the receiver operating characteristic curve (AUC) and each algorithm's advantages in the test cohort. Multimodal radiomics prediction models based on T, LPLN, and their combination were constructed, and the AUC values and predictive performance at the optimal classification threshold were compared to select the optimal radiomics prediction model. Subsequently, a clinical-radiomics combined prediction model was built based on the optimal radiomics prediction model and clinical pathological features. AUC values of the optimal radiomics prediction model and the clinical-radiomics combined prediction model were compared, along with their predictive performance at the optimal classification threshold, to select the optimal prediction model. Calibration curves and decision curve analysis (DCA) were used to evaluate the calibration and clinical effectiveness of the optimal prediction model.

Results: CatBoost was selected as the optimal classifier to build the multimodal radiomics prediction model. Compared to radiomics prediction models based solely on T or LPLN, the model based on their combination showed higher AUC, specificity (SPE), and positive predictive value (PPV) in the test cohort, making it the optimal radiomics prediction model. In the test cohort, the clinical-radiomics combined prediction model had a higher AUC value compared to the optimal radiomics prediction model (0.96 vs. 0.92, P < 0.05), with improvements in sensitivity, SPE, PPV, and negative predictive value, leading to its selection as the optimal prediction model for LPLNM. Calibration curves and DCA demonstrated that the clinical-radiomics combined prediction model exhibited good calibration and clinical effectiveness.

Conclusion: 1. EL based on multimodal MRI radiomics is feasible in predicting LPLNM in clinical practice for rectal cancer. The LPLNM prediction model constructed using a combination of clinical features and multimodal radiomic features of T and LPLN has good discrimination, calibration, and clinical effectiveness, showing good generalization ability in multicenter and prospective test cohorts. 2. EL based on multimodal MRI radiomics shows great potential in accurately predicting LPLNM in rectal cancer. 3. CatBoost has excellent performance in processing complex, high-dimensional multimodal MRI radiomics data, and can effectively deal with nonlinear relationships in the data, which helps to improve the generalization ability and accuracy of the model.