重建算法和层厚对AI提取pGGN一阶特征的影响  

The impact of reconstruction algorithm and slice thickness on the first-order features of pure ground-glass nodule extracted by artificial intelligence

作　　者：王玉凤[1] 赵青[1] 孙黎[1] 史珊[1] WANG Yufeng;ZHAO Qing;SUN Li;SHI Shan(Guang’anmen Hospital,China Academy of Chinese Medicine Sciences,Beijing 100053,China)

机构地区：[1]中国中医科学院广安门医院放射科,北京100053

出　　处：《青岛大学学报（医学版）》2024年第3期440-444,共5页Journal of Qingdao University(Medical Sciences)

基　　金：中国中医科学院广安门医院所级科研基金资助项目(2022 S467)。

摘　　要：目的 探讨不同重建算法和层厚对基于人工智能(AI)获得的纯磨玻璃结节(pGGN)一阶特征的影响,选取最佳重建算法和重建层厚。方法 回顾性分析我院同一台CT机上行胸部高分辨率平扫者CT影像,在AI辅助下由一名副主任医师筛选出最小径≥5 mm的pGGN 158例,分别记录相同重建算法不同层厚(0.625、1.250、2.500 mm)和相同层厚不同重建算法(标准算法和肺算法)数据,在AI软件中标准敏感度下获得16个一阶特征(总体积、总质量、CT最大值、CT最小值、CT平均值、CT值方差、球型度、最大面面积、表面积、3D长径、长短径平均值、峰度、偏度、能量、紧凑度、熵),并进行统计分析。结果 2.500 mm层厚数据中有7例pGGN无论标准算法还是肺算法在AI上均未识别,其余层厚和算法组合均能准确识别;另有4例pGGN在肺算法中未识别,而在标准算法及其他层厚组合中均能识别。不管标准算法还是肺算法0.625 mm组和1.250 mm组只有能量(Z=3.39、3.34,P<0.05)、熵(Z=6.49、6.77,P<0.05)差异有统计学意义;0.625 mm组和2.500 mm组在标准算法中总体积、CT最大值、CT最小值、CT平均值、表面积、球型度、紧凑度、能量、熵差异有统计学意义(Z=3.67~13.40,P<0.05),在肺算法中除上述特征外还有CT值方差、峰度差异有统计学意义(Z=2.64~13.34,P<0.05);1.250 mm组和2.500 mm组在标准算法中CT平均值、CT最大值、球型度、表面积、紧凑度、能量、熵差异有统计学意义(Z=2.43~6.98,P<0.05),在肺算法中CT最大值、CT最小值、CT值方差、球型度、表面积、紧凑度、峰度、熵差异有统计学意义(Z=2.54~8.51,P<0.05)。结论 不同的重建算法和层厚对AI获得的pGGN一阶特征均会产生影响,建议采用1.250 mm层厚、标准算法数据进行一阶特征分析。Objective To investigate the impact of reconstruction algorithm and slice thicknesses on the first-order features of pure ground-glass nodule(pGGN)extracted by artificial intelligence(AI),and to select the optimal reconstruction algorithm and slice thickness.Methods A retrospective analysis was performed for the CT images of the patients who underwent high-re-solution chest plain scan on the same CT machine in our hospital,and with the assistance of AI,158 cases of pGGN with a minimum diameter of≥5 mm were identified by an associate chief physician.Data were recorded under the conditions of the same reconstruction algorithm with different slice thicknesses(0.625,1.250,and 2.500 mm)or the same slice thickness with different reconstruction algorithms(the standard algorithm and the lung algorithm).A total of 16 first-order features were obtained under the standard sensitivity in AI software,i.e.,total volume,total mass,maximum CT value,minimum CT value,mean CT value,CT value variance,sphericity,maximum area,surface area,3D major axis,mean major and minor axes,kurtosis,skewness,energy,compactness,and entropy,and a statistical analysis was performed.Results As for the data of the slice thickness of 2.500 mm,7 cases of pGGN were not recognized by either standard algorithm or lung algorithm,while the remaining cases were accurately re-cognized on the other combinations of slice thickness and algorithm;4 cases of pGGN were not recognized by the lung algorithm,but they were recognized by the combinations of standard algorithm and other slice thicknesses.Based on both the standard algorithm and the lung algorithm,there were significant differences between the 0.625 mm group and the 1.250 mm group in energy(Z=3.39,3.34;P<0.05)and entropy(Z=6.49,6.77;P<0.05).Based on the standard algorithm,there were significant diffe-rences between the 0.625 mm group and the 2.500 mm group in total volume,maximum CT value,minimum CT value,mean CT value,surface area,sphericity,compactness,energy,and entropy(Z=3.67-13.40,P<0.05),as well as

关 键 词：磨玻璃结节 机器学习 图像处理 计算机辅助 算法 

分 类 号：R563[医药卫生—呼吸系统] R814.42[医药卫生—内科学]