| 摘要: |
| ［摘要］ 目的 基于苏木精-伊红（HE）染色切片肿瘤区域提取深度学习特征构建膀胱癌患者生存预后的预测模型。方法 收集癌症基因组图谱（TCGA）数据库379例膀胱癌患者的临床资料［包含450张全视野数字切片（WSI）］，以及暨南大学附属广东省第二人民医院2017年9月至2024年5月收治的179例膀胱癌患者的临床资料（包含244张WSI）。应用ResNet50模型进行迁移学习，识别肿瘤区域。应用RetCCL模型提取深度学习特征，通过单因素Cox回归与LASSO回归对提取的深度学习特征进行筛选，并构建风险评分模型。采用最大化选择秩统计量的方法，确定深度学习特征风险评分的最佳截断值，据此将患者分为高危组与低危组，通过Kaplan-Meier生存曲线比较两组生存预后。通过Cox回归分析影响膀胱癌患者生存预后的因素，并基于筛得风险因素指标构建列线图模型。通过校正曲线与决策曲线分析（DCA）综合评估该模型的校准度与临床净获益。结果 应用RetCCL模型对所有WSI进行特征提取（每张WSI有14 336个特征），通过单因素Cox分析得到22个有预后预测价值的特征，进一步对这22个特征进行LASSO回归分析，得到16个回归系数非0的特征。基于此16个病理深度学习特征和对应的回归系数构建膀胱癌患者的风险评分模型。针对TCGA数据库和该院临床资料的分析结果表明，基于深度学习特征风险评分，高危组的总生存期显著短于低危组（P<0.05）。多因素Cox回归分析结果显示，深度学习特征风险级别、M分期是影响膀胱癌患者生存预后的独立风险因素（P<0.05）。基于此二指标构建预测膀胱癌患者生存预后的列线图模型。校正曲线显示模型预测患者术后1年、3年和5年生存率与实际生存率之间表现出良好的一致性。DCA结果显示，在术后1年、3年和5年的生存预后方面，该预测模型决策可取得较好的临床净获益。结论 基于HE染色切片图像深度学习特征构建的膀胱癌患者生存预后预测模型具有良好的预测效能，可为临床提供精准的个体化预后评估工具。 |
| 关键词: 膀胱癌 病理 全视野数字切片 深度学习 生存预后 |
| DOI：10.3969/j.issn.1674-3806.2026.01.05 |
| 分类号:R 737.14 |
| 基金项目:国家自然科学基金项目（编号：82302304）；广东省第二人民医院博士工作站基金项目（编号：2022BSGZ011）；广东省第二人民医院托举工程基金项目（编号：TJGC-2022009）；广州市科学技术局基础研究计划基金项目（编号：2024A04J4159） |
|
| Construction of a prediction model for the survival prognosis of bladder cancer patients based on deep learning features extracted from tumor regions in hematoxylin and eosin-stained slides |
|
LUO Guanshui1, HE Yifeng2, ZHENG Zongtai3, ZENG Deqin2
|
|
1.Center for Minimally Invasive Comprehensive Cancer Therapy, Minhang Campus, the Affiliated Guangdong Second Provincial General Hospital of Jinan University, Guangzhou 510422, China; 2.General Affairs Office, Minhang Campus, the Affiliated Guangdong Second Provincial General Hospital of Jinan University, Guangzhou 510422, China; 3.Department of Urology, the Affiliated Guangdong Second Provincial General Hospital of Jinan University, Guangzhou 510317, China
|
| Abstract: |
| ［Abstract］ Objective To construct a prediction model for the survival prognosis of bladder cancer patients based on deep learning features extracted from tumor regions in hematoxylin and eosin(HE)-stained slides. Methods The clinical data of 379 patients with bladder cancer were collected from The Cancer Genome Atlas(TCGA) database［including 450 slices of whole-slide images(WSI)］, and the clinical data of 179 patients with bladder cancer who were admitted to the Affiliated Guangdong Second Provincial General Hospital of Jinan University from September 2017 to May 2024 were collected(including 244 slices of WSI). The ResNet50 model was applied for transfer learning to identify tumor regions. The RetCCL model was applied for extracting deep learning features. The extracted deep learning features were screened via univariate Cox regression and LASSO regression, and a risk scoring model was constructed. The maximally selected rank statistics method(MSRSM) was adopted to determine of the risk scores of deep learning features. According to the optimal cut-off values, the patients were divided into high-risk group and low-risk group, and the survival prognosis was compared between the two groups by using Kaplan-Meier survival curve. The factors influencing the survival prognosis of the bladder cancer patients were analyzed by using Cox regression, and a nomogram model was constructed based on the screened risk factor indicators. The calibration accuracy and net clinical benefit of the model were comprehensively evaluated by using calibration curve and decision curve analysis(DCA). Results The RetCCL model was applied to extract the features from all the WSI, and each slice of WSI had 14 336 features. Twenty-two features with prognostic predictive value were obtained by using univariate Cox regression analysis. Furthermore, LASSO regression analysis was performed on these 22 features to obtain 16 features with non-zero regression coefficients. Based on this, 16 pathological deep learning features and the corresponding regression coefficients were used to construct a risk scoring model for the bladder cancer patients. The results of analyses of TCGA database and clinical data from the Affiliated Guangdong Second Provincial General Hospital of Jinan University showed that based on the risk scores of deep learning features, the overall survival of the high-risk group was significantly shorter than that of the low-risk group(P<0.05). The results of multivariate Cox regression analysis showed that the risk level of deep learning features and M stage were independent risk factors affecting the survival prognosis of the patients with bladder cancer(P<0.05). Based on these two indicators, a nomogram model for predicting the survival prognosis of the bladder cancer patients was constructed. The calibration curve showed that the model exhibited a good consistency between the predicted survival rates of the patients at 1 year, 3 years and 5 years after surgery and the actual survival rates. The results of DCA showed that the decision-making of the prediction model could achieve good net clinical benefits in terms of survival prognosis at 1 year, 3 years and 5 years after surgery. Conclusion The survival prognosis prediction model for bladder cancer patients constructed according to deep learning features extracted from tumor regions in HE-stained slides has good predictive efficacy and can provide precise individualized prognosis assessment tools for clinical practice. |
| Key words: Bladder cancer Pathology Whole-slide images(WSI) Deep learning Survival prognosis |
