增强磁共振成像深度学习模型术前预测增殖型肝细胞癌的临床价值

刘力榛; 成杰; 陈凤喜; 李依蔓; 徐阳; 陈伟; 蔡萍; 李青锐; 李晓明

doi:10.3760/cma.j.cn115610-20250716-00475

增强磁共振成像深度学习模型术前预测增殖型肝细胞癌的临床价值

Clinical value of enhanced magnetic resonance imaging-based deep learning model in pre-operative prediction of proliferative hepatocellular carcinoma

摘要

摘要:
探讨增强磁共振成像（MRI）深度学习模型术前预测增殖型肝细胞癌的临床价值。
采用回顾性队列研究方法。收集2017年5月至2022年10月陆军军医大学第一附属医院和重庆医科大学附属第二医院收治的906例肝细胞癌患者的临床资料；男769例,女137例；年龄为（53.2±10.9）岁。906例患者中,陆军军医大学第一附属医院收治的815例患者通过随机数字表法按8∶2比例分为训练集634例和内部验证集181例,重庆医科大学附属第二医院收治的91例患者为外部验证集。训练集用于构建预测模型,验证集用于验证预测模型。观察指标：（1）肝细胞癌患者病理学分型的影响因素分析。（2）肝细胞癌患者深度学习影像特征。（3）增殖型肝细胞癌预测模型的效能评价。（4）增殖型肝细胞癌预测模型的验证。（5）肝细胞癌患者预后情况。正态分布的计量资料组间比较采用独立样本t检验,偏态分布的计量资料组间比较采用Mann⁃Whitney U检验。计数资料组间比较采用χ²检验。采用二元逻辑回归分析进行多因素分析。模型性能通过五折交叉验证评估,绘制受试者工作特征曲线,以曲线下面积（AUC）、灵敏度、特异度评价模型诊断价值。采用Delong检验进行模型诊断效能对比。采用Hosmer⁃Lemeshow检验模型校准度。通过最大约登指数确定预测模型最佳截断值,>0.175为高风险患者,≤0.175为低风险患者。采用Kaplan‑Meier法计算生存率,Log⁃rank检验进行生存分析。
（1）肝细胞癌患者病理学分型的影响因素分析。634例训练集患者中,增殖型肝细胞癌190例,非增殖型肝细胞癌444例。多因素分析结果显示：AFP水平≥400 μg/L、肿瘤最大径>5 cm是影响肝细胞癌病理学类型为增殖型的独立危险因素优势比=1.73、1.88,95%可信区间（CI）为1.19~2.50、1.30~2.71,P<0.05。（2）肝细胞癌患者深度学习影像特征。634例训练集患者中,增殖型和非增殖型肝细胞癌患者MRI深度学习预测概率分别为84.8%（30.5%,95.4%）和5.8%（3.2%,12.5%）,两者比较,差异有统计学意义（Z=-16.01,P<0.05）。（3）增殖型肝细胞癌预测模型的效能评价。训练集中,临床预测模型预测增殖型肝细胞癌的AUC为0.63（95%CI为0.59~0.68,P<0.05）,灵敏度为54.74%,特异度为64.19%；MRI深度学习预测模型预测增殖型肝细胞癌的AUC为0.90（95%CI为0.87~0.93,P<0.05）,灵敏度为80.53%,特异度为86.94%；MRI深度学习联合临床预测模型预测增殖型肝细胞癌的AUC为0.90（95%CI为0.87~0.93,P<0.05）,灵敏度为83.16%,特异度为86.04%。DeLong检验结果显示：MRI深度学习联合临床预测模型与临床预测模型比较,差异有统计学意义（P<0.05）,与MRI深度学习预测模型比较,差异无统计学意义（P>0.05）。Hosmer‑Lemeshow检验结果显示：临床预测模型、MRI深度学习预测模型、MRI深度学习联合临床预测模型拟合优度均较好（χ²=0.84、6.38、3.93,P>0.05）。这提示3种预测模型的预测概率与实际风险匹配程度较高。（4）增殖型肝细胞癌预测模型的验证。内部验证集验证结果显示：MRI深度学习预测模型预测增殖型肝细胞癌的AUC为0.84（95%CI为0.77~0.91,P<0.05）,灵敏度为82.35%,特异度为77.69%；外部验证集验证结果显示：MRI深度学习预测模型预测增殖型肝细胞癌的AUC为0.81（95%CI为0.71~0.92,P<0.05）,灵敏度为70.00%,特异度为81.69%。（5）肝细胞癌患者预后情况。906例患者中,645例增殖型肝细胞癌患者1、3、5年无复发生存率分别为56.9%、31.4%、29.1%,261例非增殖型肝细胞癌患者1、3、5年无复发生存率分别为88.8%、68.6%、56.0%。训练集、内部验证集、外部验证集增殖型肝细胞癌患者无复发生存期与非增殖型肝细胞癌患者比较,差异均有统计学意义（P<0.05）。331例高风险肝细胞癌患者1、3、5年无复发生存率分别为64.6%、50.4%、43.6%,575例低风险肝细胞癌患者1、3、5年无复发生存率分别为88.5%、71.9%、62.7%。训练集、内部验证集、外部验证集高风险肝细胞癌患者无复发生存期与低风险肝细胞癌患者比较,差异均有统计学意义（P<0.05）。
MRI深度学习模型术前可良好预测增殖型肝细胞癌及患者无复发生存期。

Abstract:
Objective To investigate the clinical value of enhanced magnetic resonance imaging (MRI)-based deep learning model in preoperative prediction of proliferative hepatocellular carcinoma (HCC).
Methods The retrospective cohort study was conducted. The clinical data of 906 HCC patients who were admitted to The First Affiliated Hospital of Army Medical University and The Second Affiliated Hospital of Chongqing Medical University from May 2017 to October 2022 were collected. There were 769 males and 137 females, aged (53.2±10.9)years. Of the 906 patients, 815 cases who were admitted to The First Affiliated Hospital of Army Medical University were divided into the training set of 634 patients and the internal validation set of 181 patients using a random number table method with a ratio of 8:2, and 91 patients who were admitted to The Second Affiliated Hospital of Chongqing Medical University were divided into the external validation set. The training set was used to construct the prediction model, while the validation set was used to validate the prediction model. Observation indicators: (1) analysis of factors influencing the pathological classification of HCC patients; (2) deep learning imaging features of HCC patients; (3) evaluation of the efficacy of prediction model for proliferative HCC; (4) validation of the prediction model for proliferative HCC; (5) prognosis of HCC patients. Comparison of measurement data with normal distribution between groups was conducted using the independent sample t test. Comparison of measurement data with skewed distribution between groups was conducted using the Mann-Whitney U test. Comparison of count data between groups was conducted using the chi-square test. Multivariate analysis was conducted using the binary Logistic regression model. The model perfor-mance was evaluated through five-fold cross-validation, and receiver operating characteristic (ROC) curve was plotted to assess the diagnostic value of the model based on the area under curve (AUC), sensitivity, and specificity. The Delong test was used to compare the diagnostic performance of models. The Hosmer-Lemeshow test was employed to evaluate the calibration of models. The optimal cutoff value of the prediction model was determined by the maximum Youden index, with the value >0.175 indicating high-risk patients and value ≤0.175 indicating low-risk patients.The Kaplan‑Meier method was used to calculate the survival rate and the Log-rank test was used for survival analysis.
Results (1) Analysis of factors influencing the pathological classification of HCC patients. Of 634 patients in the training set, there were 190 cases of proliferative HCC and 444 cases of non-proliferative HCC. Results of multivariate analysis showed that alpha fetoprotein (AFP) ≥400 μg/L and tumor diameter >5 cm were independent risk factors for pathological type of HCC as proli-ferative odds ratio=1.73, 1.88, 95% confidence interval (CI) as 1.19-2.50, 1.30-2.71, P<0.05. (2) Deep learning imaging features of HCC patients. In the training set of 634 patients, the probability predicted by MRI-based deep learning model was 84.8%(30.5%,95.4%) for proliferative HCC and 5.8%(3.2%,12.5%) for non-proliferative HCC, showing a significant difference between them (Z=-16.01, P<0.05). (3) Evaluation of the efficacy of prediction model for proliferative HCC. In the training set, the AUC of clinical prediction model for proliferative HCC was 0.63(95%CI as 0.59-0.68, P<0.05), with sensitivity of 54.74% and specificity of 64.19%. The AUC of MRI-based deep learning prediction model was 0.90(95%CI as 0.87-0.93, P<0.05), with sensitivity of 80.53% and specificity of 86.94%. The AUC of combined MRI-based deep learning with clinical prediction model was 0.90 (95%CI as 0.87-0.93, P<0.05), with sensitivity of 83.16% and specificity of 86.04%. Results of Delong test showed that there was a significant difference between the combined MRI-based deep learning with clinical prediction model and the clinical prediction model (P<0.05), and there was no signifi-cant difference between the combined MRI-based deep learning with clinical prediction model and the MRI-based deep learning prediction model (P>0.05). Results of Hosmer-Lemeshow test showed good calibration for the clinical prediction model, the MRI-based deep learning prediction model and the combined MRI-based deep learning with clinical prediction model (χ²=0.84, 6.38, 3.93, P>0.05), indicating that the predicted probabilities of these three prediction models matched the actual risk well. (4) Validation of the prediction model for proliferative HCC. Results of validation of the prediction model in internal validation set showed the AUC of MRI-based deep learning prediction model for proliferative HCC was 0.84(95%CI as 0.77-0.91, P<0.05), with sensitivity of 82.35% and specificity of 77.69%. Results of validation of the prediction model in external validation set showed the AUC of MRI-based deep learning prediction model for proliferative HCC was 0.81(95%CI as 0.71-0.92, P<0.05), with sensitivity of 70.00% and specificity of 81.69%. (5) Prognosis of HCC patients. Of the 906 patients, the 1-, 3-, and 5-year recurrence-free survival rates for 645 proliferative HCC patients were 56.9%, 31.4%, and 29.1%, respectively, and the 1-, 3-, and 5-year recurrence-free survival rates for 261 non-proliferative HCC patients were 88.8%, 68.6%, and 56.0%, respectively. There were significant differences in recurrence-free survival time between proliferative HCC and non-proliferative HCC patients of the training set, internal validation set and external validation set (P<0.05). The 1-, 3-, 5-year recurrence-free survival rates for 331 high-risk HCC patients were 64.6%, 50.4%, 43.6%, versus 88.5%, 71.9%, 62.7% for 575 low-risk HCC patients. There were significant differences in recurrence-free survival time between high-risk HCC patients and low-risk HCC patients of the training set, internal validation set and external validation set (P<0.05).
Conclusion The MRI-based deep learning model can effectively predict proliferative HCC and recurrence-free survival of patients before the surgery.

HTML全文

参考文献(27)

施引文献

资源附件(0)