首页|期刊导航|分子影像学杂志|基于ResNet101技术的多模态磁共振成像2.5D深度学习模型可有效预测肝细胞癌细胞角蛋白-19表达

基于ResNet101技术的多模态磁共振成像2.5D深度学习模型可有效预测肝细胞癌细胞角蛋白-19表达OA

A novel 2.5D deep learning model integrated with ResNet101 based on multimodal magnetic resonance imaging enables effective prediction of CK-19 expression in hepatocellular carcinoma

中文摘要英文摘要

目的 构建并验证基于Resenet101的多模态MRI 2.5D深度学习(DL)模型预测肝细胞癌(HCC)细胞角蛋白-19(CK-19)表达.方法 回顾性分析2018年1月~2025年1月在苏州大学附属第三医院接受术前MRI的218例HCC的患者,按照7∶3的比例随机分为训练集(n=152)、测试集(n=66).依据免疫组化结果将CK-19归类为阳性、阴性表达.基于ResNet101训练DL模型,提取T2加权成像(T2WI)、动脉期(AP)、门静脉期(PVP)的2.5D定量特征.采用LASSO回归筛选最佳特征后输入极端梯度提升算法构建多模态MRI 2.5D模型.采用ROC曲线下面积(AUC)、校准曲线、决策分析曲线评估模型诊断CK-19表达的效能、一致性及临床受益.应用梯度加权类激活映射(Grad-CAM)生成HCC内部可视化热力图.结果 T2WI 2.5D模型诊断CK-19阳性表达的训练集AUC值为0.81(95%CI:0.74~0.88)、测试集为0.71(95%CI:0.54~0.86),AP 2.5D模型为0.83(95%CI:0.76~0.89)、0.74(95%CI:0.69~0.84),PVP 2.5D模型为0.80(95%CI:0.72~0.88)、0.70(95%CI:0.57~0.88).MRI融合模型诊断效能最高,AUC值分别为0.87(95%CI:0.80~0.93)、0.78(95%CI:0.64~0.93).校准曲线和决策分析曲线显示MRI 2.5D模型预测CK-19表达具有较高的稳定性和临床净受益.Grad-CAM实现可视化HCC内部CK-19阳性表达区域.结论 基于ResNet101的多模态MRI 2.5D模型可有效预测CK-19表达,尤其是MRI融合模型具有较高稳定性及临床净受益.

Objective To develop and validate a 2.5D deep learning(DL)model integrated with the ResNet101 architecture based on multimodal MRI for the preoperative prediction of cytokeratin-19(CK-19)expression in hepatocellular carcinoma(HCC).Methods A total of 218 patients diagnosed with HCC who underwent preoperative MRI at the Third Affiliated Hospital of Soochow University from January 2018 to January 2025 were retrospectively enrolled,and randomly divided into a training cohort(TC,n=152)and a validation cohort(VC,n=66)at a ratio of 7:3.CK-19 expression was determined as either positive or negative according to immunohistochemical analysis.A 2.5D DL model was constructed on the ResNet101 framework,and quantitative imaging features were extracted from T2-weighted imaging(T2WI),arterial phase(AP),and portal venous phase(PVP)MRI sequences.The LASSO algorithm was applied to screen the most robust and discriminative imaging features,which were subsequently incorporated into a 2.5D DL model with an Extreme Gradient Boosting classifier.The predictive performance of each model for CK-19 expression was assessed using the AUC.Calibration curves and decision curve analysis(DCA)were utilized to evaluate the predictive consistency and clinical net benefit of the established models.Gradient-weighted Class Activation Mapping(Grad-CAM)was performed to visualize the intuitive heat maps within HCC lesions.Results For the single-sequence models,the T2WI-based 2.5D model achieved AUC values of 0.81(95%CI:0.74-0.88)in the TC and 0.71(95%CI:0.54-0.86)in the VC.The AP-based 2.5D model yielded AUCs of 0.83(95%CI:0.76-0.89)and 0.74(95%CI:0.69-0.84)in the TC and VC,respectively.The PVP-based 2.5D model exhibited AUC values of 0.80(95%CI:0.72-0.88)for the TC and 0.70(95%CI:0.57-0.88)for the VC.Notably,the multimodal combined 2.5D MRI model obtained the optimal predictive performance,with AUCs of 0.87(95%CI:0.80-0.93)in the TC and 0.78(95%CI:0.64-0.93)in the VC,accompanied by favorable calibration and superior clinical net benefit.Grad-CAM visualization successfully localized the focal regions corresponding to CK-19 positive expression in HCC tissues.Conclusion The ResNet101-based 2.5D model derived from multimodal MRI enables effective prediction of CK-19 expression in HCC.In particular,the integrated multimodal MRI model achieves high predictive stability and favorable clinical net benefit.

秦莉;刘海峰

苏州大学附属第三医院/常州市第一人民医院医学影像科,江苏 常州 213000苏州大学附属第三医院/常州市第一人民医院医学影像科,江苏 常州 213000

磁共振成像ResNet 101深度学习肝细胞癌细胞角蛋白-19

magnetic resonance imagingResNet 101deep learninghepatocellular carcinomacytokeratin-19

《分子影像学杂志》 2026 (8)

1051-1058,8

江苏省常州市科技局项目(CJ20244017)苏州大学附属第三医院临床研究专项(2024-14)苏州大学附属第三医院临床应用新技术(2026-12)

10.12122/j.issn.1674-4500.2026.08.10

评论