目的：评估人工智能辅助胃癌肝转移病灶自动分割软件对胃癌肝转移病灶的分割功能相对于手动分割,能否减少观察者内及观察者间的差异。方法：由2名医生盲法应用西门子医疗开发的基于深度学习网络的肝脏肿瘤分析软件（Liver Lesion Analysis Tool）,分别使用全手动模式以及人工智能辅助的半自动模式,对32例患者共81个胃癌肝转移灶的CT图像进行分割,并对最长径及三维体积进行纯手动和全自动重复测量。应用Bland-Altman法分别评估最长径测量及体积测量在2种分割模式下的观察者内及观察者间差异,用组内相关系数（intraclass correlation coefficient, ICC）评估2种模式下最长径测量与体积测量的观察者内及观察者间测量变异度差异。结果：手动、半自动模式最长径测量的观察者内95%一致性区间分别为-31.70%~34.55%,-28.04%~27.89%,手动、半自动模式最长径测量的观察者间95%一致性区间分别为-74.26%~38.85%,-59.54%~40.98%。手动、半自动模式体积测量的观察者内95%一致性区间分别为-148.10%~102.70%,-75.92%~63.79%,手动、半自动模式体积测量的观察者间95%一致性区间分别为-127.40%~111.50%,-87.66%~43.77%。观察者内手动模式最长径测量与体积测量变异度的ICC分别为0.914、0.950,观察者内半自动模式最长径测量与体积测量变异度的ICC分别为0.967、0.970。观察者间手动模式最长径测量与体积测量变异度的ICC分别为0.884、0.939,观察者间半自动模式最长径测量与体积测量变异度的ICC分别为0.928、0.974。相对于手动分割病灶,使用人工智能辅助的肝转移自动分割软件进行半自动病灶分割所需要的时间大大缩短[（25.78±4.23） s比（4.53±2.82） s, P<0.01]。结论：使用人工智能辅助的肝转移自动分割软件进行半自动病灶分割在观察者内及观察者间的一致性好,可提高肝转移病灶分割的效率,有望成为临床随访及疗效评估的定量工具。

Objective: To evaluate whether the differences within intra-observer and inter-observer can be reduced by the assisted segmentation based on artificial intelligence algorithm fordetection of liver metastasis from gastric cancer. Methods: Thirty two patients with liver metastasis from gastric cancer were retrospectively analyzed. Two radiologists blindly applied the Liver Lesion Analysis Tool (Siemens Healthineers, not for commercial use) for performing repeated measurement of the longest diameter and three-dimensional volume of liver lesions in CT images in a semi-automatic mode and by the manual mode. The measurements were repeated by one of the two radiologist after two weeks. The Bland-Altman method was used to evaluate the intra-observer and inter-observer differences. The intra-group correlation coefficient(ICC) was used to estimate the difference in variability. Results: In intra-observer analysis, the 95% consistency interval of the manual and semi-automatic longest diameter measurements were -31.70% -34.55% and -28.04% -27.89%, respectively. The 95% consistency interval ininter-observer of manual and semi-automatic longest diameter measurements were -74.26% -38.85% and -59.54% -40.98%, respectively. The 95% consistency interval inintra-observer of the manual and semi-automatic volume measurements were -148.10% -102.70% and -75.92% -63.79%, respectively, and in inter-observer were -127.40% -111.50% and -87.66% -43.77%, respectively. The ICC of the manual longest diameter measurement and the volume measurement variability inintra-observer were 0.914 and 0.950, respectively (P<0.001), and those in semi-automatic mode were 0.967 and 0.970, respectively (P<0.001). The ICC of the longest diameter and the volume measurement variability in inter-observer were 0.884 and 0.939, respectively (P<0.001), and those in semi-automatic were 0.928 and 0.974, respectively (P<0.001). The time required for semi-automatic mode was significantly shorter than that required for manual mode[（25.78±4.23） s vs （4.53±2.82） s, P<0.01]. Conclusions: The semi-automatic segmentation of liver metastasis using artificial intelligence-assisted automatic segmentation software is reproducible and can improve the efficiency of segmentation of liver metastases. The AI-assisted liver analysis tool has the potential to be a quantitative tool for clinical follow-up and treatment evaluation.