【摘要】 目的 利用深度学习（deeplearning,DL）结合超声造影（contrast-enhancedultrasound,CEUS）视频开发一个时空混合感知的DL模型，从肝脏影像报告与数据系统（liverimagingreportinganddatasystem,LI-RADS）LR-M类肝结节中无创诊断肝细胞癌（hepatocellularcarcinoma,HCC）。方法 对来自中山大学附属第一医院和广西医科大学第一附属医院的经CEUS诊断为LR-M类的患者进行回顾性分析，以病理和综合诊断为标准，分为HCC组和非HCC组。使用R(2+1)D-18卷积神经网络对CEUS视频进行多尺度时空分时相并行特征提取，构建HCC诊断模型，并通过帧采样策略（15帧、30帧、45帧）平衡信息完整性与计算效率，辅以梯度加权分类激活映射技术（gradient-weightedclassactivation mapping,Grad-CAM）可视化解释决策依据。使用准确率、敏感度、特异度、阳性预测值、阴性预测值和受试者工作特征（receiveroperatingcharacteristic,ROC）曲线来评估模型的诊断效能。结果 来自中山大学附属第一医院的176例LR-M患者按7∶3划分训练集和验证集；来自广西医科大学第一附属医院的46例LR-M患者为测试集。当以45帧视频作为输入时，模型诊断效能最佳，在验证集和测试集中，敏感度为91.67%和84.21%，特异度为96.43%和81.48%，曲线下面积（area undercurve,AUC）为0.940和0.828。与甲胎蛋白（alpha-fetoprotein,AFP）指标相比，模型对HCC的检出率更高（验证集：50.00%，测试集：31.58%;P值均<0.05），联合使用DL模型与AFP可以优化整体诊断效能。结论 基于CEUS时空感知的深度学习模型可以精准诊断LR-M类中的HCC，诊断效能显著优于AFP，两者联合诊断有望减少不必要的肝穿刺活检。更多还原

【Abstract】 Objective To develop a spatiotemporal hybrid perception deep learning(DL) model using contrast-enhanced ultrasound(CEUS) videos for diagnosis of hepatocellular carcinoma(HCC) in liver imaging reporting and data system(LI-RADS)category LR-M liver nodules. Methods This study conducted a retrospective analysis of patients diagnosed with LR-M nodules by CEUS from the First Affiliated Hospital of Sun Yat-sen University and the First Affiliated Hospital of Guangxi Medical University.Using pathological and comprehensive diagnostic results as the gold standard, the patients were categorized into HCC and non-HCC.A multi-scale spatiotemporal split-phase parallel feature extraction of CEUS video using R(2+1)D-18 convolutional neural network is used to construct a diagnostic model for HCC, and the information integrity and computational efficiency are balanced by a frame sampling strategy(15/30/45 frames), supplemented by gradient-weighted class activation mapping(Grad-CAM) visualisation to explain the decision basis. The diagnostic efficacy of the model is evaluated using accuracy, sensitivity, specificity, positive predictive value, negative predictive value and receiver operating characteristic(ROC) curve. Results A total of 176 LR-M patients from the First Affiliated Hospital of Sun Yat-sen University were divided into training and validation sets in a 7:3 ratio, while 46 LR-M patients from the First Affiliated Hospital of Guangxi Medical University served as the test set. When using 45-frame videos as input, with sensitivities of 91.67% and 84.21%, specificities of 96.43% and 81.48%, and area under curves(AUCs) of 0.940 and 0.828 in the validation and test sets. The model had a higher detection rate of HCC compared with the alpha-fetoprotein(AFP) index(validation set: 50.00%, test set: 31.58%; both P value < 0.05), and the combined use of the DL model with AFP could optimize the overall diagnostic efficacy. Conclusion The DL model with CEUS spatiotemporal perception enables accurate diagnosis of HCC in LR-M category liver lesions, demonstrating significantly superior performance over AFP. Moreover, the combined diagnosis of the both can reduce unnecessary liver biopsy.更多还原