Value of radiomics in differential diagnosis of chromophobe renal cell carcinoma and renal oncocytoma

In 2004, the World Health Organization formally classified chRCC as a new pathological classification of renal tumors. The incidence of chRCC is second only to that of renal clear cell and renal papillary cell carcinomas [3]; moreover, it has potential for metastasis. RO is a benign tumor with good prognosis [5]. Presently, surgical treatment, including partial and radical nephrectomy, is an effective method for treating local renal tumors. Radical nephrectomy can lead to an increased risk of chronic kidney disease, and is associated with an increased risk of cardiovascular disease morbidity and mortality. Compared to radical nephrectomy, a partial nephrectomy can preserve partial renal function, reduce overall mortality, and reduce the incidence of cardiovascular disease [2]. Therefore, radical nephrectomy should be avoided when nephron retention is achievable. Percutaneous renal biopsy is the most commonly used preoperative examination method, and it has 97% accuracy rate for distinguishing malignant renal masses [15]. However, the diagnosis of chRCC and RO on percutaneous renal biopsy presents difficulties [16].

chRCC and RO have many overlapping imaging features [17]. Clinically, it is difficult to distinguish chRCC from RO only by visual imaging. Currently, controversies exist on the imaging manifestations of the two kinds of tumors. For example, Rosenkrantz [10] stated that MRI features including fat, hemorrhage, margin of the mass, perirenal fat infiltration, renal vein cancer thrombus, enhancement uniformity, vascular proliferation and central scar, and segmental enhancement inversion cannot be used to distinguish between chRCC and RO. Wu et al. [9] reported that cases of RO present more instances of central scar, radial enhancement, and segmental enhancement inversion than those observed for chRCC on contrast-enhanced CT. Kim et al. [18] asserted that differential diagnosis between different tumor types can be achieved by CT enhancement. Most cases of chRCC showed homogeneous enhancement, whereas most renal clear cell and papillary cell carcinomas showed heterogeneous enhancement. Additionally, RO is characterized by homogeneous enhancement of the solid mass. The enhancement of RO and chRCC at each stage is lower than that of the normal renal cortex; however, enhancements of RO are more prominent than those of chRCC [9]. Thus, in clinical practice, it is difficult to distinguish chRCC from RO only by visual imaging. Therefore, this study used radiomics to differentiate chRCC from RO.

Texture analysis refers to the process of extracting texture feature parameters through certain image processing technologies, so as to obtain quantitative or qualitative description of texture. This technique can be used to detect subtle differences that cannot be detected by the naked eye and is more objective for tumor discrimination. Because chRCC and RO are relatively rare compared to renal clear cell and renal papillary cell carcinoma, radiomic studies of renal tumors are focused on relatively common renal tumors. Studies on the most frequently occurring renal clear cell carcinoma have focused on different aspects such as preoperative diagnosis [19‐22], tumor grade [23], prognostic evaluation [24], and molecular analysis of the cancer genes [25‐27]. Yu et al. [20] extracted the texture features of four types of renal tumors, including renal clear cell carcinoma, renal papillary cell carcinoma, chRCC, and RO. The tumors were classified by an SVM classifier, and the histogram feature median demonstrated an AUC of 0.882 for differentiating chRCC from RO. In this study, SVM was used to classify the features screened by CMP and NP. The AUC of differential diagnosis between chRCC and RO was found to be as high as 0.964, which is better than previously reported results. Zhang et al. [19] combined several texture features including SD, entropy, mean positive pixels, and kurtosis to differentiate renal clear cell carcinoma from non-transparent cell carcinoma. The value of the AUC was 0.94 ± 0.03 and the accuracy was 0.87 (sensitivity = 89%, specificity = 92%). Similar methods were used to differentiate between renal papillary cell carcinoma and chRCC, and the accuracy of differential diagnosis was 78%. Most of the subjects in the above study had renal malignant tumors, but no benign ROs were included in the comparison. Thus, this study has more clinical significance in terms of comparing chRCC with RO.

Considering the isodensity of tumor masses on plain CT scans, errors may easily occur while describing an ROI. Therefore, this study analyzed the contrast-enhanced images; however, we did not analyze the CT plain scan images. Some studies such as the one by Hodgdon et al. [21] only analyzed CT plain scan images of renal tumors. Using CT plain scan texture features and subjective visual features to differentiate and diagnose fat-deficient angiomyolipoma from other renal tumors, the accuracy of the classification methods based on texture features was found to be higher than that of radiologists’ subjective judgment of tumors or that observed with the use of an SVM classifier to identify renal tumors. The accuracy of differential diagnosis was about 83–91%. Schieda et al. [28] did not analyze the vascular characteristics of renal masses according to the nuclear grading system for chRCC; therefore, only CT plain scan images were used for radiomic analysis for clinical grading of the tumors. Importantly, to ensure the accuracy of the boundary delineation of the CT plain scans, it is still necessary to use the image of CT enhancement as a reference. Kocak et al. [23] analyzed the influence of different edge segmentation methods on feature selection and classification performance, including contour focusing and edge contraction by 2 mm. The results show that the latter method can extract more texture features; however, the former method has better reproducibility of features and better classification performance for the nuclear grading system-based classification of renal clear cell carcinoma. In this study, manual ROI extraction was performed to segment the edge contour of each transverse mass. Recently, a variety of mathematical techniques have been used in radiomics to quantify image textures, including statistical, Fourier, and wavelet analysis, and have been applied to the study of a variety of tumors. Varghese et al. [22] used multi-phase CT fast Fourier transform index to analyze the CT-enhanced solid and fat-deficient renal masses. Good classification results were obtained when distinguishing benign from the malignant renal masses, differentiating RO from chRCC, and RO from lipid-poor angiomyolipoma (AUC > 0.7).

Bektas et al. [29] used different machine learning classifiers, such as SVM, MLP, RF, kNN, and naive Bayes, for predicting Fuhrman nuclear grade of clear cell renal cell carcinomas (ccRCCs), and the best model was created using SVM (AUC = 0.851, accuracy = 0.913). Lee et al. [30] combined different feature selection methods and different feature classifiers, which included SVM, RF, kNN and LR, to distinguish benign fat-poor angiomyolipoma from malignant ccRCC. kNN and SVM classifiers with ReliefF feature selection achieved the best accuracy of 72.3 ± 4.6% and 72.1 ± 4.2%, respectively. The results of this study show that five classifiers have good diagnostic performance in feature classification methods (accuracy > 0.89, AUC > 0.90). One of the best models for the differential diagnosis between chRCC and RO was the use of SVM to classify the features screened by the combination of CMP and NP. The accuracy was found to be as high as 0.945, and the AUC was 0.964 ± 0.054. We suggest the use of radiomics of enhanced CT images for differentiating between chRCC and RO.

This study has some limitations. First, this was a single-center study and the sample size was small; notably, there were relatively fewer RO cases. A multi-center study of these rare cases must be undertaken under favorable conditions. Second, the CT equipment was not uniform and the scanning parameters were different, which may have influenced the repeatability of the results. Third, failure to analyze plain CT images may have led to the omission of internal masses. Considering the isodensity of tumor masses on plain CT scans, errors could have occurred when describing the ROIs for some characteristic information. To address this issue, only prominent enhancement phase images were analyzed.

In summary, we established a machine learning model that can distinguish chRCC from RO on enhanced CT images. These models are expected to help clinicians formulate better clinical diagnosis and devise improved treatment strategies. Our results indicate that radiomics can accelerate the development of personalized therapy.