第一部分：PSMA PET/CT联合双参数MRI对前列腺癌根治术后病理升级和不良病理特征的预测作用

目的：本研究旨在探讨基于临床资料、双参数磁共振（Magnetic Resonance Imaging，MRI）和前列腺特异性膜抗原正电子发射断层扫描/计算机断层扫描（Prostate-specific Membrane Antigen Positron Emission Tomography/Computed Tomography，PSMA PET/CT）影像组学特征的多模态模型对前列腺癌靶向穿刺病理升级、前列腺包膜外侵犯（Extra-Prostatic Extension，EPE）和高级别前列腺癌的预测价值。

方法：回顾性纳入本中心根治性前列腺切除手术（Radical Prostatectomy，RP）术前接受PSMA PET/CT和双参数MRI检查的前列腺癌患者的临床和影像资料。根据前列腺病理大切片确定每位患者靶向穿刺病理升级（升级组和未升级组）、病理分期（≥pT3和pT2）、病理分级[高级别（ISUP分组≥3组）和低级别（ISUP分组≤2组）]的二分类结局。将患者按照7：3随机分为训练集和测试集，分别在PET序列、MRI的T2WI和ADC序列上进行图像分割。针对靶向穿刺病理升级，感兴趣区域（Volume of Interest，VOI）为主要病灶，即病理大切片中Gleason评分最高的病灶；针对病理分期，VOI为前列腺腺体；针对病理分级，VOI为前列腺癌病灶。使用Pyradiomics工具包提取影像组学特征，对特征进行降维后，使用逻辑回归分别构建PET和MRI影像组学模型，单因素和多因素逻辑回归构建临床模型。最终通过不同的模态组合构建四种融合模型。使用受试者工作特征曲线下面积（Area under the Receiver Operating Characteristic Curve，ROC-AUC）、准确性、灵敏度、特异度、精确度和F1值评估模型预测效能。采用Delong检验比较模型之间的差异。

结果：共有117、122和122例患者用于构建靶向穿刺病理升级、病理分期和病理分级的影像组学多模态预测模型。穿刺ISUP分组是靶向穿刺病理升级的独立预测因子（ISUP分组1组 OR=1.57，95%CI 1.25-1.97），基于临床、MRI和PET影像组学特征构建的融合模型在预测前列腺癌靶向穿刺病理升级的效能最佳（测试集AUC=0.899，95%CI 0.779-1.000，准确性=0.861），显著优于临床模型（P=0.006）。PSA[PSA≥20ng/ml（OR=1.79，95%CI 1.25-2.55）]和临床分期（OR=1.31，95%CI 1.01-1.70）是EPE的独立危险因素，联合MRI和临床因素的融合模型在鉴别EPE的效能最优（AUC=0.839，95%CI 0.688-0.989，准确性=0.757），而增加PET影像组学特征后模型效能下降。另外，前列腺影像报告和数据系统（Prostate Imaging Reporting and Data System，PI-RADS）评分（OR=1.22，95%CI 1.07-1.39）与高级别前列腺癌相关，融合临床、MRI和PET的影像组学模型可更好地识别高级别前列腺癌（AUC=0.917，95%CI 0.825-1.000，准确性=0.865）。

结论：联合临床特征、双参数MRI和PSMA PET的多模态影像组学模型能够准确预测前列腺癌的靶向穿刺病理升级和高级别病理。在鉴别EPE方面，基于临床特征和双参数MRI的影像组学模型表现最佳。

第二部分：基于病理组学和机器学习在鉴别病理大切片中前列腺癌不良病理特征的价值研究

目的：本研究旨在探讨基于病理组学特征和机器学习模型在预测前列腺病理大切片中高级别前列腺癌、筛状结构和磷酸酶和张力蛋白同源物（Phosphatase and Tensin Homologue, PTEN）蛋白表达缺失的价值。

方法：回顾性纳入本中心接受RP手术的前列腺癌患者，通过病理大切片确定每位患者病理分级[高级别（ISUP分组≥3组）和低级别（ISUP分组≤2组）]、筛状结构（有和无）、PTEN表达情况（表达和缺失）的二分类结果。按照7：3随机分为训练集和测试集，利用CellProfiler图像分析软件确定细胞和细胞核轮廓，并提取细胞、细胞核及病理图像三个层面的病理组学特征。采用统计检验、相关性分析和LASSO回归进行病理组学特征的降维，将保留的特征纳入到八种机器学习算法中，包括逻辑回归、朴素贝叶斯、支持向量机、随机森林、极端随机树、极端梯度提升、轻量梯度提升机和多层感知器，分别构建病理组学模型。使用ROC-AUC、准确性、灵敏度、特异度、精确度和F1值评估模型预测效能，最终挑选出预测不同不良病理特征的最佳模型。

结果：本研究共纳入329例患者，在954张切片中勾画肿瘤病灶，共切割图像块285034张，其中有202例患者用于预测PTEN表达情况。高级别组共有100例（30.4%），基于逻辑回归构建的病理组学模型在识别高级别前列腺癌的效能最佳，且在训练集（AUC=0.926，95%CI 0.892-0.960，准确性=0.865）和测试集（AUC=0.826，95%CI 0.736-0.916，准确性=0.828）中性能稳定。筛状结构组共有199例（60.5%），基于随机森林算法构建的病理组学模型在鉴别前列腺癌筛状结构的效能最佳（训练集AUC=0.934，95%CI 0.900-0.967，准确性=0.861；测试集AUC=0.853，95%CI 0.776-0.929，准确性=0.808）。然而，对于PTEN表达缺失组，通过八种机器学习算法构建的病理组学模型在预测PTEN表达的效能均不佳（测试集AUC<0.7，准确性<0.7）。

结论：本研究建立了基于病理组学特征的高级别前列腺癌和筛状结构的识别模型，为病理学家的诊断提供有价值的参考。

第三部分：基于病理和影像深度学习对前列腺癌根治术后生化复发的预测作用

目的：本研究旨在探讨基于临床资料、双参数MRI和病理深度学习特征的不同组合，开发预测前列腺癌RP术后生化复发（Biochemical Recurrence，BCR）的最优融合模型。

方法：回顾性纳入本中心自2017年8月至2022年8月确诊前列腺癌并接受根治手术的患者，所有患者均有病理大切片和术前MRI资料。BCR被定义为连续两次随访 PSA 值回升至0.2ng/ml以上且有上升趋势。将患者按照7：3随机分为训练集和测试集，分别在病理大切片、MRI的T2WI和ADC序列上进行图像分割，使用预训练的ResNet-50深度学习模型，提取病理和影像深度学习特征，通过多示例学习将图像块水平预测概率转换为全切片图像（Whole Slide Imaging，WSI）水平的特征。采用单因素COX回归分析、相关性分析和LASSO回归进行深度学习特征的降维，将保留的特征纳入到多因素COX回归中分别构建病理和影像深度学习模型。临床模型包括基于单、多因素COX回归构建的多因素临床模型、CAPRA（Cancer of the Prostate Risk Assessment）模型和CAPRA-S（the CAPRA Post-surgical Score）模型。最后，通过对不同的模态的组合，使用C指数（Concordance Index）和时间依赖ROC-AUC评价模型的效能，从中挑选出预测RP术后BCR的最佳融合模型。采用KM曲线（Kaplan-Meier Curve）和log-rank检验评估不同分组的预后情况。

结果：本研究共纳入220例前列腺癌患者，有43例（19.5%）发生BCR，中位随访时间为37个月。多因素COX回归分析结果显示PSA（HR=1.03，95%CI 1.00-1.06，P=0.042）和RP术后ISUP分组（HR=1.52，95%CI 1.09-2.14，P=0.015）是BCR的独立危险因素。联合病理深度学习评分和CAPRA-S评分的术后融合模型，在预测RP术后BCR方面表现最佳（C指数=0.829，95%CI 0.717-0.940），优于其他单一模态模型，包括多因素临床模型（C指数=0.659，95%CI 0.460-0.857）、CAPRA模型（C指数=0.660，95%CI 0.507-0.814）、CAPRA-S模型（C指数=0.724，95%CI 0.583-0.864）和病理深度学习模型（C指数=0.771，95%CI 0.665-0.877）。在术后融合模型中增加影像深度学习评分并未提高融合模型的预测效能（C指数=0.807，95%CI 0.696-0.919）。KM曲线显示，基于术后融合模型评分的高低风险分组可有效区分患者的预后（P=0.022）。

结论：联合病理深度学习评分和CAPRA-S评分构建的术后融合模型可用于预测RP术后前列腺癌BCR的风险，其预测效能优于单一模态的临床、影像和病理模型，该模型有助于识别可能从辅助治疗中受益的患者。

Part I: The Predictive Role of PSMA PET/CT Combined with Biparametric MRI in the Pathological Upgrading and Adverse Pathological Features of Prostate Cancer after Radical Prostatectomy

Objective: This study aims to investigate the predictive value of a multimodal model based on clinical data, biparametric magnetic resonance imaging (MRI), and prostate-specific membrane antigen positron emission tomography/computed tomography (PSMA PET/CT) radiomic features for MRI-targeted biopsy pathological upgrading, extra-prostatic extension (EPE), and high-grade prostate cancer (PCa).

Methods: Clinical and imaging data of patients with PCa who underwent radical prostatectomy (RP) and preoperative PSMA PET/CT and biparametric MRI examinations were retrospectively collected. Pathological upgrading (upgraded vs. non-upgraded), pathological staging (≥pT3 vs. pT2), and pathological grading [high-grade (ISUP grade group ≥3) vs. low-grade (ISUP grade group ≤2)] were determined based on prostate pathology slides. Patients were randomly divided into training and testing sets (7:3 ratio). Image segmentation was performed on PET sequences, T2-weighted imaging (T2WI), and apparent diffusion coefficient (ADC) sequences of MRI. The volume of interest (VOI) for pathological upgrading was defined as the primary lesion with the highest Gleason score on pathology slides. Pyradiomics toolkit was used to extract radiomic features. After feature selection, logistic regression models were built for PET, MRI radiomics separately. Clinical models were constructed using univariable and multivariable logistic regression. Four fusion models were developed by combining different modalities. Model performance was evaluated using the area under the receiver operating characteristic curve (ROC-AUC), accuracy, sensitivity, specificity, precision, and F1 score. Differences between models were compared using the Delong test.

Results: A total of 117, 122, and 122 patients were included for the development of radiomics multimodal predictive models for MRI-targeted biopsy pathological upgrading, staging, and grading, respectively. ISUP grade group was found to be an independent predictor of pathological upgrading (ISUP grade group 1: OR=1.57, 95% CI 1.25-1.97). The fusion model based on clinical, MRI, and PET radiomic features demonstrated the best performance in predicting pathological upgrading of PCa (testing set AUC=0.899, 95% CI 0.779-1.000, accuracy=0.861), significantly outperforming the clinical model (P=0.006). PSA [PSA ≥20 ng/ml (OR=1.79, 95% CI 1.25-2.55)] and clinical staging (OR=1.31, 95% CI 1.01-1.70) were independent risk factors for EPE. The fusion model combining MRI and clinical factors showed the best performance in distinguishing EPE (AUC=0.839, 95% CI 0.688-0.989, accuracy=0.757), while adding PET radiomic features resulted in decreased model performance. Moreover, Prostate Imaging Reporting and Data System (PI-RADS) score (OR=1.22, 95% CI 1.07-1.39) was associated with high-grade PCa. The fusion radiomics model combining clinical, MRI, and PET features could better identify high-grade PCa (AUC=0.917, 95% CI 0.825-1.000, accuracy=0.865).

Conclusion: The multimodal radiomics model combining clinical features, biparametric MRI, and PSMA PET/CT accurately predicts pathological upgrading and high-grade PCa in patients with PCa. For identifying EPE, the model based on clinical features and biparametric MRI performs the best.

Part II: Investigation of the Value of Pathomics and Machine Learning in Identifying Adverse Pathological Features of Prostate Cancer in Pathological Whole Slide Images

Objective: This study aims to explore the value of pathomics features combined with machine learning models in predicting high-grade PCa, cribriform architecture, and loss of Phosphatase and Tensin Homologue (PTEN) protein expression in PCa.

Methods: In our retrospective study, we enrolled PCa patients who had undergone RP at our center. We categorized each patient's pathological grade (high-grade: ISUP grade group ≥ 3 and low-grade: ISUP grade group ≤ 2), presence of cribriform architecture (present or absent), and PTEN expression status (expressed or lost) based on pathological whole slide images (WSI). Patients were randomly divided into training and testing sets at a 7:3 ratio. Utilizing CellProfiler image analysis software, we identified cell and nuclear contours and extracted pathomics features from three levels: cellular, nuclear, and pathological image. Feature reduction was performed using statistical tests, correlation analysis, and LASSO regression, and the retained features were incorporated into eight machine learning algorithms: logistic regression, naive Bayes, support vector machine, random forest, extremely randomized trees, extreme gradient boosting, light gradient boosting machine, and multilayer perceptron, to construct pathomics models. Model performance was evaluated using ROC-AUC, accuracy, sensitivity, specificity, precision, and F1 score, and the best model for predicting various adverse pathological features was selected.

Results: A total of 329 patients were included, with 954 slides delineating tumor lesions, generating 285,034 image patches. Among these, 202 patients were utilized to predict PTEN expression status. There were 100 cases (30.4%) in the high-grade group, with the logistic regression-based pathomics model demonstrating optimal performance in identifying high-grade prostate cancer, showing stable performance in both training (AUC=0.926, 95% CI 0.892-0.960, accuracy=0.865) and testing sets (AUC=0.826, 95% CI 0.736-0.916, accuracy=0.828). In the cribriform architecture group, comprising 199 cases (60.5%), the pathomics model developed with the random forest algorithm showcased optimal performance in identifying cribriform architecture in PCa. Specifically, in the training set, the AUC was 0.934 (95% CI 0.900-0.967) with an accuracy of 0.861, while in the testing set, the AUC was 0.853 (95% CI 0.776-0.929) with an accuracy of 0.808. However, for the PTEN expression loss group, the pathomics models built using eight machine learning algorithms showed poor performance in predicting PTEN expression (testing set AUC<0.7, accuracy<0.7).

Conclusion: This study developed identification models for high-grade PCa and cribriform architecture using pathomics features, providing valuable references for pathologists' diagnostic processes.

Part III: Prediction of Biochemical Recurrence in Prostate Cancer After Radical Prostatectomy Based on Pathological and MRI Deep Learning

Objective: This study aims to develop an optimal fusion model for predicting biochemical recurrence (BCR) after RP in PCa, utilizing different combinations of clinical data, biparametric MRI, and pathological deep learning features.

Methods: A retrospective cohort of patients diagnosed with PCa and undergoing RP surgery at our center from August 2017 to August 2022 was included. All patients had pathological whole-mount slides and preoperative MRI data. BCR was defined as two consecutive PSA values rising above 0.2 ng/ml with an upward trend. Patients were randomly divided into training and testing sets in a 7:3 ratio. Image segmentation was performed on pathological whole-mount slides and MRI sequences including T2WI and ADC. A pre-trained ResNet-50 deep learning model was used to extract pathological and MRI deep learning features. Multiple-instance learning was employed to convert image-level prediction probabilities into whole-slide imaging (WSI)-level features. Univariable Cox regression analysis, correlation analysis, and LASSO regression were used for feature dimensionality reduction, followed by constructing pathological and MRI deep learning models using multivariable Cox regression. Clinical models included multivariable clinical models based on multivariable Cox regression, the Cancer of the Prostate Risk Assessment (CAPRA) model, and the CAPRA-S (the CAPRA Post-surgical Score) model. Finally, the performance of different model combinations was evaluated using concordance index (C-index) and time-dependent ROC-AUC, and the best fusion model for predicting post-RP BCR was selected. Kaplan-Meier curves and log-rank tests were used to evaluate the prognosis of different groups.

Results: A total of 220 PCa patients were included, among whom 43 (19.5%) experienced BCR, with a median follow-up time of 37 months. Multivariable Cox regression analysis showed that PSA (HR=1.03, 95% CI 1.00-1.06, P=0.042) and post-RP ISUP grade group (HR=1.52, 95% CI 1.09-2.14, P=0.015) were independent risk factors for BCR. The postoperative fusion model combining pathological deep learning scores and CAPRA-S scores performed best in predicting post-RP BCR (C-index=0.829, 95% CI 0.717-0.940), outperforming other single-modal models, including multi-factor clinical model (C-index=0.659, 95% CI 0.460-0.857), CAPRA model (C-index=0.660, 95% CI 0.507-0.814), CAPRA-S model (C-index=0.724, 95% CI 0.583-0.864), and pathological deep learning model (C-index=0.771, 95% CI 0.665-0.877). Adding MRI deep learning scores to the postoperative fusion model did not improve its predictive performance (C-index=0.807, 95% CI 0.696-0.919). Kaplan-Meier curves showed that risk stratification based on postoperative fusion model scores effectively distinguished patients' prognosis (P=0.022).

Conclusion: A postoperative fusion model combining pathological deep learning scores and CAPRA-S scores can be used to predict the risk of post-RP BCR in PCa with better predictive performance than single-modal clinical, MRI, and pathological models. This model can help identify patients who may benefit from adjuvant therapy.