CT放射组学对肾乏脂肪血管平滑肌脂肪瘤的预测价值

Predictive value of CT-based radiomics in renal angiomyolipomas without visible fat

目的肾乏脂肪血管平滑肌脂肪瘤(AMLwvf)是一种良性肿瘤,预后良好,其影像表现与恶性肿瘤相似。本研究通过放射组学方法从每位患者四期CT图像中提取有意义特征数据,并建立机器学习模型从常见肾脏肿瘤中区分出AMLwvf。方法回顾性分析29例病理证实AMLwvf与169例其他常见肾肿瘤病例。所有病例CT图像病灶内均未发现肉眼可见脂肪,均行CT平扫(PCP)及皮髓质期(CMP)、肾实质期(NP)、排泄期(EP)增强扫描。首先对每位患者各期图像中病灶行手工绘制容积感兴趣区(VOI),提取放射组学特征,为降低模型冗余度,采用最小绝对收缩和选择算子用于特征筛选,最后采用五折交叉验证的机器学习分类器进行鉴别诊断,具体包括K最近邻、逻辑回归、多层感知器、支持向量机。分类器的性能主要通过受试者工作曲线的曲线下面积(AUC)和准确率、敏感度、特异度来评价和比较。结果对每个病灶每期VOI所提取特征进行筛选,最终PCP、CMP、NP、EP以及CMP+NP等CT放射组学的特征分别为10、10、14、2、7个。其中CMP和NP在使用多层感知器后达到了令人满意的性能,AUC分别为0.82±0.06和0.85±0.06,准确率分别为83.33%(敏感度为0.6896,特异度为0.8579)和75.76%(敏感度为0.7241,特异度为0.7633)。EP和PCP的AUC分别为0.73±0.16和0.79±0.06,准确率分别为78.79%(敏感度为0.6207,特异度为0.8166)和79.29%(敏感度为0.5862,特异度为0.8284),将CMP和NP结合后,AUC为0.86±0.06,准确率为70.20%(敏感度为0.8621,特异度为0.6746)。结论基于放射组学特征的机器学习能区分良性的AMLwvf与常见肾脏肿块,这有助于临床对肾肿瘤患者干预方式的选择。

Objective Renal angiomyolipomas without visible fat(AMLwvf)are benign tumors with a good prognosis.The imaging findings of AMLwvf are similar to those of renal malignant tumors.In this study,radiomics was used to extract meaningful characteristic data from four phases of CT images of each patient,and a machine learning model was established to distinguish AMLwvf from common renal tumors.Methods This retrospective study included 29 patients with AMLwvf confirmed by pathology and 169 patients with other common renal tumors.No visible fat was seen in the lesions on all CT images.Non-contrast-enhanced CT[precontrast phase(PCP)]and contrast-enhanced CT[corticomedullary phase(CMP),nephrographic phase(NP),and excretory phase(EP)]scanning were performed on all patients.The volume region of interest(VOI)was manually drawn from images of each phase of each patient to extract radiomics features.In order to reduce the redundancy of the model,the least absolute shrinkage and selection operator was used for feature selection.Differential diagnosis was done by machine learning classifiers with 5-fold cross-validation,including k-nearest neighbor,logistic regression,multilayer perception,and support vector machine.The performance of classifiers was mainly evaluated and compared by the area under ROC curve(AUC),accuracy,sensitivity and specificity.Results The extracted features of VOI of each lesion were screened,and the final CT radiomics features of PCP,CMP,NP,EP and CMP+NP were 10,10,14,2 and 7,respectively.Among them,CMP and NP achieved satisfactory performance after using multilayer perception,with the AUC of 0.82±0.06 and 0.85±0.06,respectively,and the accuracy of 83.33%(sensitivity 0.6896;specificity 0.8579)and 75.76%(sensitivity 0.7241;specificity 0.7633),respectively.The AUC of EP and PCP was 0.73±0.16 and 0.79±0.06,respectively,and the accuracy was 78.79%(sensitivity 0.6207;specificity 0.8166)and 79.29%(sensitivity 0.5862;specificity 0.8284),respectively.In the combination of CMP and NP,the AUC was 0.86±0.06 and the accuracy was 70.20%(sensitivity 0.8621;specificity 0.6746).Conclusion Machine learning based on radiomics features can distinguish benign AMLwvf from common kidney tumors,which may contribute to the selection of clinical interventions for patients with renal tumors.