Department of Radiology, Zhejiang Provincial People's Hospital, People's Hospital of Hangzhou Medical College, 310000, Hangzhou, China.
Department of Neurosurgery, Shaoxing City Keqiao District Hospital of Traditional Chinese Medicine, 312000, Shaoxing, China.
Abdom Radiol (NY). 2020 Aug;45(8):2500-2507. doi: 10.1007/s00261-020-02414-9.
This study aimed to discriminate fat-poor angiomyolipoma (fp-AML) from clear cell renal cell carcinoma (ccRCC) by constructing radiomics-based logistic classifiers in comparison with conventional computed tomography (CT) analysis at three CT phases.
Twenty-two fp-AML patients and 62 ccRCC patients who were pathologically identified were enrolled in this study, and underwent three-phase (unenhanced phase, UP; corticomedullary phase, CMP; nephrographic phase, NP) CT examinations. Whole-tumor regions of interest (ROIs) were contoured in ITK software by two radiologists. Radiomic features were dimensionally reduced by means of ANOVA + MW, correlation analysis, and LASSO. Four radiomics logistic classifiers including the UP group, CMP group, NP group, and sum group were built, and the radiomic scores (rad-scores) were calculated. After collecting the qualitative and quantitative conventional CT characteristics, the conventional CT analysis logistic classifier and the radiomics-based logistic classifier were constructed. The receiver operating characteristic curve (ROC) was constructed to evaluate the validity of each classifier.
The area under curve (AUC) of the conventional CT analysis logistic classifier including angular interface, cyst degeneration, and pseudocapsule was 0.935 (95% CI 0.860-0.977). Regarding logistic classifiers for radiomics analysis, the AUCs of the UP group were 0.950 (95% CI 0.895-1.000) and 0.917 (95% CI 0.801-1.000) in the training set and testing set, respectively, which were higher than those of the CMP and NP groups. The AUCs of the sum group were observed to be the highest. The top three selected features for the UP and sum groups belonged to GLCM parameters and histogram parameters. The radiomics-based logistic classifier encompassed cyst degeneration, pseudocapsule, and sum rad-score, and the AUC of the model was 0.988 (95% CI 0.935-1.000).
Whole-tumor radiomics-based CT analysis is superior to conventional CT analysis in the differentiation of fp-AML from ccRCC. Cyst degeneration, pseudocapsule, and sum rad-score are the most significant factors. The radiomics analysis of the UP group shows a higher AUC than that of the CMP and NP groups.
本研究旨在通过构建基于放射组学的逻辑分类器，并与三期 CT 分析比较，从乏脂性血管平滑肌脂肪瘤(fp-AML)和透明细胞肾细胞癌(ccRCC)中进行鉴别。
本研究纳入了 22 例经病理证实的 fp-AML 患者和 62 例 ccRCC 患者，所有患者均行三期 CT 检查（平扫期、皮质期、肾盂期）。由两位放射科医生在 ITK 软件中勾画全肿瘤感兴趣区(ROI)。采用方差分析+MW、相关性分析和 LASSO 对放射组学特征进行降维。构建了基于 UP 组、CMP 组、NP 组和总和组的四个放射组学逻辑分类器，并计算了放射组学评分(rad-scores)。在收集定性和定量的常规 CT 特征后，构建了常规 CT 分析逻辑分类器和基于放射组学的逻辑分类器。构建受试者工作特征曲线(ROC)评估每个分类器的有效性。
包含角状界面、囊肿变性和假包膜的常规 CT 分析逻辑分类器的曲线下面积(AUC)为 0.935(95%CI 0.860-0.977)。在基于放射组学的逻辑分类器方面，训练集和测试集 UP 组的 AUC 分别为 0.950(95%CI 0.895-1.000)和 0.917(95%CI 0.801-1.000)，高于 CMP 和 NP 组。总和组的 AUC 最高。UP 组和总和组的前三个特征分别属于 GLCM 参数和直方图参数。基于放射组学的逻辑分类器包含囊肿变性、假包膜和总和 rad-score，模型的 AUC 为 0.988(95%CI 0.935-1.000)。
与常规 CT 分析相比，基于全肿瘤放射组学的 CT 分析在 fp-AML 与 ccRCC 的鉴别诊断中具有优势。囊肿变性、假包膜和总和 rad-score 是最重要的因素。与 CMP 和 NP 组相比，UP 组的放射组学分析 AUC 更高。
