不同CT扫描设备对人工智能深度学习模型 定量测定结果的对比研究

目的 评估不同CT扫描设备对基于深度学习算法模型定量测定结果的准确性和稳定性影响。方法 制作225例不同体积、密度的标准容积水球，采用分层抽样法分为训练集（n=45）、测试集（n=180），训练集用于模型的建立，测试集用于测试模型的准确性。采用GE Revolution 256排512层CT和Siemens SOMATOM Defintion AS 64排128层螺旋CT扫描获得影像资料，以标准容积水球作为验证标准，比较模型测量不同CT扫描设备来源数据的测量准确性及稳定性。结果 将两种不同扫描设备来源的数据用同一个模型进行体积测量比较，GE Revolution 256排512层 CT和Siemens SOMATOM Defintion AS 64排128层螺旋CT的百分误差分别为2.050和7.837，差异有统计学意义（P<0.001），变异系数分别为0.029和0.055，差异有统计学意义（P<0.05）。结论 基于深度学习的全自动智能体积测量模型具有较高的准确性，且测量准确性、稳定性和一致性均受扫描设备的影响。

Comparative study on quantitative measurement results of artificial intelligence deep learning model by different CT scanning devices

Objective To evaluate the influence of different CT scanning devices on the accuracy and stability of quantitative determination results based on deep learning algorithm model. Methods A total of 225 cases of standard volume water balloons of different volumes and densities were made, which were divided into the training set (n=45) and the test set (n=180) according to stratified sampling method. The training set was used for model building and the test set was used to test the accuracy of the model. Image data were obtained by GE Revolution 256-row 512-slice CT and Siemens SOMATOM Defintion AS 64-row 128-slice spiral CT scanning, and standard volume water balloons were used as verification criteria. The model was compared to measure the accuracy and stability of data from different CT scanning devices. Results Compared volumetric measurements from two different scanning device sources using the same model, the percent error was 2.050 and 7.837 for the GE Revolution 256-row 512-slice CT and Siemens SOMATOM Defintion AS 64-row 128-slice spiral CT, with statistically significant differences (P<0.001) and coefficients of variation of 0.029 and 0.055, with statistically significant differences(P<0.05). Conclusion The automatic intelligent volume measurement model based on deep learning has high accuracy, and the measurement accuracy, stability and consistency are affected by the scanning equipment, but it needs to be verified by a large sample of clinical trials in the future.