4 DISCUSSION
CCSK, with unclear histological and immunohistological features, is an unusual pediatric renal malignancy secondary to Wilms’ tumor.4 The main clinical manifestations of CCSK include abdominal mass, abdominal pain, hematuria and other atypical symptoms similar to Wilms’ tumor, and the lack of specificity of these symptoms can result in misdiagnosis of CCSK as Wilms’ tumor preoperatively.3 Thus, the accurate preoperative differentiation between CCSK and Wilms’ tumor is helpful for clinical decision making in children. Kang et al found perinephric vessel engorgement and higher tumor enhancement on CT images are useful in differentiating CCSK from Wilms’ tumor.7 However, these qualitative and semi-quantitative features cannot reflect intra-tumoral heterogeneity comprehensively. Recently, radiomics provides a promising method to evaluate tumor phenotypes quantitatively.12,13 In this study, we used CT-based radiomics analysis to differentiate CCSK from Wilms’ tumor for the first time. Some differences were confirmed in radiomics features between CCSK and Wilms’ tumor.
According to our results, among the final optimal features screened through the variance threshold, SelectKBest and LASSO methods, all of them were first-order features depicting the distribution of voxel intensities, of which skewness from NP images achieved moderate to good diagnostic performance for CCSK. A skewness is about the asymmetry of the distribution of voxel intensities,14 indicating that there are differences in the asymmetry of the voxel intensity histogram between CCSK and Wilms’ tumor. Additionally, the performance of skewness transformed by exponential and squareroot filters is obviously superior to that of original skewness, suggesting advanced features filtered by filters could reveal more invisible meaningful information about tumoral heterogeneity.15 In previous literatures, first-order histogram characteristics of renal tumors differed in various pathological types.10,16,17 Deng et al investigated the role of CT texture analysis in differentiating major renal cell carcinoma subtypes, and the first-order entropy was found to be the most meaningful biomarker in differentiating clear cell from papillary renal neoplasms.17 Likewise, skewness and kurtosis were demonstrated to be helpful for differentiating clear cell renal carcinoma from oncocytoma.16 However, in other studies on texture analysis of non-renal tumors, second-order features, such as gray level size zone matrix or gray level difference matrix, seem to play a more important role in characterizing heterogeneity of non-renal tumors.18-20
When training with LR model, all the selected features from each phase were used to construct classification model for diagnosing CCSK. The results showed LR models combining all chosen features perform better than the majority of single feature. Because it is hard to delineate the boundary of tumor lesion from kidney on non-contrast-enhanced CT images, we only chose CMP and NP images to perform radiomics analysis. Compared to CMP images, NP images provided more useful data to the CCSK-associated radiomics characteristics. When combining CMP and NP images, an interesting finding was that all the optimal features are extracted from the NP images. Meanwhile, the performance of the composite model was similar to that of NP model, suggesting two-phase CT images have no additional value in differentiating between CCSK and Wilms’ tumor. Meng et al demonstrated NP features are the most sensitive features for characterizing sarcomatoid from clear cell renal carcinoma, the reason for which may be that sarcomatoid differentiation causes changes in intra-tumoral enhancement patterns.11 Boo et al found there are some unique vascular patterns in CCSK, in which regularly-spaced fibrovascular septa separates the nests of tumor cells, and this may cause late enhancement in CCSK compared with Wilms’ tumor.21
Despite LR model based on NP images had moderate to good performance in diagnosing CCSK, the results of Delong test showed no significant difference between LR model, exponential-skewness and squareroot-skewness based on NP in training and validation set, which further confirms the important role of skewness in differentiating CCSK from Wilms’ tumor. The distribution of exponential-skewness and squareroot-skewness in CCSK was different from Wilms’ tumor. And higher skewness was statistically associated with CCSK, and lower skewness with Wilms’ tumor. Although NP-based skewness performed moderate to good in our study, this feature may be a supplementary biomarker contributing to the differential diagnosis between CCSK and Wilms’ tumor. However, the exact utility of skewness in the CCSK and Wilms’ tumor differentiation still needs further investigation.
Admittedly, there were some limitations in the present study. First, due to the rarity of CCSK, as a tertiary referral children’s medical center, only 29 pediatric patients with CCSK were enrolled in this study. Second, considering the predominant prevalence of Wilms’ tumor, 51 patients with Wilms’ tumor were selected consecutively as control group, which may cause selection bias to our results. Third, the slice thickness of CT images in our study was 5 mm in order to include more patients as possible, and thin slice thickness may help to reflect more meaningful radiomics features between CCSK and Wilms’ tumor. Finally, the CT scans performed in our study were obtained on two different scanners. Although preprocessing of the images was performed, the radiomics features derived from different scanners may have some influence on the diagnostic performance for CCSK.
In conclusion, radiomics is a promising method to differentiate CCSK from Wilms’ tumor in children. Skewness from NP images at exponential and squareroot filters was able to discriminate between CCSK and Wilms’ tumor, obtaining moderate to good diagnostic performance for CCSK. And higher skewness on NP images may be a potential biomarker for diagnosing CCSK from Wilms’ tumor.