Department of Radiology, Mayo Clinic Hospital, 5777 E Mayo Blvd, Phoenix, AZ, 85054, USA.
School of Computing, Informatics, Decision Systems Engineering, Arizona State University, Tempe, AZ, USA.
Abdom Radiol (NY). 2018 Jun;43(6):1439-1445. doi: 10.1007/s00261-017-1331-0.
We aimed to determine the best algorithms for renal stone composition characterization using rapid kV-switching single-source dual-energy computed tomography (rsDECT) and a multiparametric approach after dataset expansion and refinement of variables.
rsDECT scans (80 and 140 kVp) were performed on 38 ex vivo 5- to 10-mm renal stones composed of uric acid (UA; n = 21), struvite (STR; n = 5), cystine (CYS; n = 5), and calcium oxalate monohydrate (COM; n = 7). Measurements were obtained for 17 variables: mean Hounsfield units (HU) at 11 monochromatic keV levels, effective Z, 2 iodine-water material basis pairs, and 3 mean monochromatic keV ratios (40/140, 70/120, 70/140). Analysis included using 5 multiparametric algorithms: Support Vector Machine, RandomTree, Artificial Neural Network, Naïve Bayes Tree, and Decision Tree (C4.5).
Separating UA from non-UA stones was 100% accurate using multiple methods. For non-UA stones, using a 70-keV mean cutoff value of 694 HU had 100% accuracy for distinguishing COM from non-COM (CYS, STR) stones. The best result for distinguishing all 3 non-UA subtypes was obtained using RandomTree (15/17, 88%).
For stones 5 mm or larger, multiple methods can distinguish UA from non-UA and COM from non-COM stones with 100% accuracy. Thus, the choice for analysis is per the user's preference. The best model for separating all three non-UA subtypes was 88% accurate, although with considerable individual overlap between CYS and STR stones. Larger, more diverse datasets, including in vivo data and technical improvements in material separation, may offer more guidance in distinguishing non-UA stone subtypes in the clinical setting.
我们旨在使用快速 kV 切换单源双能 CT(rsDECT)和多参数方法，在数据集扩展和变量细化后，确定用于肾结石成分特征描述的最佳算法。
对 38 个 5-10mm 的离体肾结石（尿酸(UA)结石 21 个，鸟粪石(STR)结石 5 个，胱氨酸(CYS)结石 5 个，一水合草酸钙(COM)结石 7 个）进行 rsDECT 扫描（80 和 140 kVp）。对 17 个变量进行测量：11 个单能 keV 水平的平均 CT 值(HU)、有效 Z 值、2 个碘水材料基本对、3 个平均单能 keV 比值(40/140、70/120、70/140)。分析包括使用 5 个多参数算法：支持向量机、随机树、人工神经网络、朴素贝叶斯树和决策树(C4.5)。
使用多种方法，UA 与非 UA 结石的分离准确率达到 100%。对于非 UA 结石，使用 70keV 平均截断值 694HU 可准确区分 COM 与非 COM（CYS、STR）结石。随机树获得区分所有 3 种非 UA 亚型的最佳结果(15/17，88%)。
对于 5mm 或更大的结石，多种方法可以 100%准确地区分 UA 与非 UA 结石以及 COM 与非 COM 结石。因此，分析方法的选择取决于用户的偏好。用于区分所有三种非 UA 亚型的最佳模型准确率为 88%，尽管 CYS 和 STR 结石之间存在相当大的个体重叠。更大、更多样化的数据集，包括体内数据和物质分离技术的改进，可能会在临床环境中更有助于区分非 UA 结石亚型。
