MR Imaging Radiomics Analyses for Prediction of High Grade Histology and Necrosis in Clear Cell Renal Cell Carcinoma: Preliminary Experience

Abstract

Introduction:

Percutaneous renal mass biopsies can accurately diagnose clear cell renal cell carcinoma (ccRCC), however their reliability to determine nuclear grade in larger, heterogeneous tumors is limited. To assess the ability of radiomics analyses of magnetic resonance imaging (MRI) to predict high grade (HG) histology in ccRCC.

Patients and Methods:

70 patients with a renal mass underwent 3T MRI before surgery between 8/2012 and 8/2017. Tumor length, first order statistics, and Haralick texture features were calculated on T2-weighted (T2W) and dynamic contrast enhanced (DCE) MRI after manual tumor segmentation. After variable clustering algorithm was applied, tumor length, wash-out and all cluster features were evaluated univariably by receiver operating characteristic (ROC) curves. Three logistic regression models were constructed to assess predictability of HG ccRCC and cross-validated.

Results:

At univariate analysis, area under the curve (AUC) of length, DCE texture cluster 1 and cluster 3 for diagnosis of HG ccRCC were 0.7 (95% CI, 0.58-0.82, false discovery rate (FDR) p-value = 0.008), 0.72 (95% confidence interval (CI), 0.59-0.84, FDR p-value = 0.004) and 0.75 (95% CI, 0.63-0.87, FDR p-value = 0.0009), respectively. At multivariable analysis, AUC for model 1 (tumor length only), model 2 (length + DCE clusters 3 and 4), and model 3 (DCE cluster 1 and 3) for diagnosis of HG ccRCC were 0.67 (95% CI, 0.54-0.79), 0.82 (95% CI, 0.71-0.92), and 0.81 (95% CI, 0.70-0.91), respectively.

Conclusion:

Radiomics analysis of MRI images was superior to tumor size for the prediction of high-grade histology in ccRCC in our cohort.

MicroAbstract:

Radiomics analyses including histogram data and Haralick texture features of magnetic resonance imaging (MRI) offer a reasonable and superior diagnostic performance compared to tumor size for the determination of tumor grade in patients with clear cell renal cell carcinoma (ccRCC). MRI-based radiomics may play an adjunct role to percutaneous renal biopsy in management decisions of ccRCC patients with heterogeneous tumors.

Introduction:

Active surveillance (AS) is contemplated by both the American Urological Association (AUA) guidelines ¹ and the European Society for Medical Oncology ² as an option for the management of T1a/T1b renal masses (i.e., tumors up to 7 cm in size limited to the kidney), particularly for patients with comorbidities and/or increased surgical risk. The use of AS is however almost exclusively reserved to small renal masses (i.e., <4cm), despite evidence indicating low risk of developing metastatic disease for patients with larger renal cell carcinomas (RCC) . Patient selection for AS is likely influenced by the known association between the risk of metastases, size, and histologic grade ^{3, 4}. Indeed, the risk of high grade histology increases with tumor diameter: only 12% of cT1a renal tumors (≤4 cm) compared to 28% of T1b tumors (4-7 cm) and 52% of T2 tumors (>7 cm) ⁵. Thus, tumor size remains as a main clinical tool to predict the likelihood of aggressive histology non-invasively and guide management. Nomograms based on age, gender, symptoms and tumor size for the prediction of high-grade tumors have been, for the most part, unreliable ^{6, 7}.

Among RCC subtypes, clear cell RCC (ccRCC) is the most common and aggressive form, encompassing approximately 70-80% of all RCCs ^{8, 9}. Therefore, a reliable diagnosis of high grade ccRCC could facilitate the adoption of AS in a larger number of renal tumors. While percutaneous renal mass biopsies (RMBs) can accurately diagnose ccRCC, their reliability to determine nuclear grade in larger, heterogeneous tumors is limited ^{10, 11}, and likely related to the genetic and pathologic heterogeneity that characterizes this disease ^{12, 13}. For example, areas of low grade and high grade histology frequently coexist in the same tumor in larger ccRCCs ¹³. Consequently, 16% of tumors undergoing RMB are upgraded from low-grade to high-grade on final surgical pathology ¹⁴. Similarly, other histopathologic features that correlate with worse outcomes in ccRCC such as tumor necrosis are not uniformly distributed in most tumors¹⁵. Thus, a diagnostic method that can predict poor prognostic histopathologic features in heterogeneous renal masses is lacking.

Multiparametric magnetic resonance (mpMR) imaging protocols including T2-weighted (T2W), T1-weighted chemical shift (i.e. in-phase [IP]/opposed-phase [OP]) and multiphasic contrast-enhanced imaging are routinely used in clinical practice for evaluation of renal masses ^16–18. Existing evidence indicates that the common subtypes of RCC and other common benign epithelial renal neoplasms (e.g. angiomyolipoma, oncocytoma) can be differentiated non-invasively with mpMR ^18–21. A recently proposed clear cell likelihood score (ccLS) based on a pre-defined diagnostic algorithm for interpretation of mpMR can predict clear cell histology in cT1a renal masses ²². Nevertheless, the current implementation of the ccLS does not allow for prediction of high grade histology in ccRCCs.

To date, proposed approaches for renal mass characterization using mpMR imaging are based on a subjective assessment of the renal mass imaging characteristics (e.g. morphology, signal intensity) and/or basic quantitative measures, usually limited to degree of contrast enhancement^{18–21, 23, 24}. More recently, radiomics or the “conversion of images to higher dimensional data and subsequent mining of these data for improved decision support” ²⁵ has gained interest in oncology as an option to generate more objective measures of disease state that are less vulnerable to variability among interpreters. A number of approaches including analyses of histogram data and texture features (e.g. Haralick features) have been successfully applied to a variety of tumors for prediction of tumor histopathologic characteristics and prognosis ^26–29.

To the best of our knowledge, the usefulness of radiomics for prediction of high grade histology in ccRCC has not been reported. The purpose of this work was to assess the ability of radiomics to predict high grade (HG) histology in patients with ccRCC evaluated with magnetic resonance imaging (MRI) prior to surgery.

Patients and Methods

Patients

The institutional review board approved this Health Insurance Portability and Accountability Act (HIPAA)-compliant prospective study. Written informed consent was obtained from all patients. Inclusion criteria were: >18 years of age, known renal mass >2.5 cm scheduled for surgical resection, willingness to undergo a research MRI, and confirmation of ccRCC diagnosis after surgery. Exclusion criteria were: contraindication for MRI, renal insufficiency (estimated glomerular filtration rate < 30 ml/min/1.73 m²), and any prior therapeutic intervention for other neoplasms. 36 of the 102 patients have been previously reported ^{30, 31} . These prior articles focused on the validation of arterial spin labelled and dynamic contrast enhanced (DCE) MRI as surrogate markers of angiogenesis ³¹ and DWI as biomarker of tumor cellularity in ccRCC ³⁰, respectively, whereas in this manuscript we report the utility of radiomics for prediction of high grade histology and necrosis in ccRCC.

Imaging protocol

All MRI examinations were performed on a 3T whole-body MRI system (Achieva or Ingenia, Philips Healthcare, Best, The Netherlands) using an anterior and posterior phased-array torso coils. MRI protocol (Table 1) included coronal and axial T2W half-Fourier single-shot turbo spin-echo (SShTSE) and coronal 3-dimensional (3D) T1-weighted spoiled gradient-echo (SPGR) with 3 different flip angles (FA: 10°, 5°, and 2°) to generate T1 maps. The same coronal DCE SPGR acquisition with a FA of 10° was obtained before, during, and after an I.V. injection of 0.1 mmol/kg of body weight of gadobutrol (Gadavist; Bayer Healthcare Pharmaceuticals, Wayne, NJ) using a power injector at a rate of 2 mL/second followed by a 20 mL saline flush at 2 mL/s. Injection of contrast and DCE MRI acquisition were initiated simultaneously and the patient received instructions to hold his/her breath at 15 seconds. After those initial 15 seconds of free breathing, 3 dynamic phases (5 sec temporal resolution) were acquired during a 15-second breath-hold followed by a 15-second period of free breathing without image acquisition. In order to minimize the respiratory motion during the acquisition, this cycle was repeated during a total acquisition time of 5 min 45 sec.

Table 1:

MRI protocol

Acquisition	TR/TE (ms)	Slice thickness/gap (mm)	Flip angle (degrees)	FOV (mm2)	Matrix
Coronal T2-weighted half-Fourier SShTSE	1115/80	5/1	90°	402 x 340	284 x 268
Axial T2-weighted half-Fourier SShTSE	1087/80	5/0	90°	400 x 320	308 x 200
Coronal 3D T1-weighted SPGR (for T1 map)	3/1.53	5/NA	10°, 5°, and 2°	180 x 408	120 x 288
Coronal 3D SPGR sequence (DCE-MRI)*	3/1.53	5/NA	10°	180 x 408	120 x 288

2D, two dimensional; 3D, three dimensional; SShTSE, single-shot turbo spin-echo; TE, echo time; TR, repetition time; SPGR, spoiled gradient-echo; NA, not applicable.

^*Total acquisition time of 5 minutes and 45 seconds at a 5 sec temporal resolution.

Image analysis

All images were transferred to open-source Digital Imaging and Communications in Medicine (DICOM) viewer (OsiriX MD) ³². Tumor size was measured on T2W images in 3 planes. Tumor length (i.e., largest dimension) was recorded. A body MRI fellowship-trained radiologist (IP, 17-year experience), who was unaware of tumor grade results at histopathology, manually segmented all tumors. A free-hand region of interest (ROI) was drawn to include the entire tumor avoiding the edge of the lesion to minimize partial volume effects using a representative center-slice on T2W images. DCE images were processed using a commercial software (VersaVue, iCAD Inc., Nashua, NH) to generate quantitative maps of Ktrans and Kep from the extended Tofts model. Significant motion was noted between different dynamic acquisitions limiting the utility of pharmacokinetic reconstructions with the Tofts model. Thus, 4 phases from the DCE MRI acquisition were selected. mimicking the timing of our multi-phase acquisition for contrast-enhanced MRI in clinical practice. Selection of these image acquisitions on DCE MRI acquisition was based on adequate enhancement and absence of motion artifacts as follows (times relative to time of contrast administration [t=0]): pre-contrast (PRE, at 5 or 10 sec), corticomedullary phase (CM, at 40-60 sec), early nephrographic (NG, at 120-150 sec), and late nephrographic/delayed (DEL, at 220-245 sec). Within each of the selected image acquisitions, we chose a center slice through the tumor for further analysis.

Normalization and First-order Statistics

Histogram analysis and first order statistics were calculated after pixel-by-pixel data extraction from all ROIs using an in-house OsiriX plugin to generate pixel co-ordinates and signal intensity (SI) values. All DCE MRI post-contrast pixel values were normalized as the percent change relative to the same tumor pre-contrast median ROI value following the equation:

$${\tilde{s}}_i=\frac{s_i-m}{m}\times 100\%$$

where ${\tilde{s}}_i$ is the normalized SI of pixel i, s_i is the original SI of pixel i and m is the median SI of the ROI on the pre-contrast phase. This approach was used to allow broad applicability of the analysis even in patients with underlying renal disease (i.e. normalization using median signal of background renal parenchyma).

Mean, standard deviation, 75^th quantile (Q3), 90^th quantile (P90), kurtosis and skewness of the ROIs were calculated on normalized DCE MRI histogram data. Peak enhancement on DCE MRI was calculated as [(SI_max − SI_pre)/SI_pre] x 100%, where SI_max is the maximum mean SI in any of the 3 post-contrast phases and SI_pre is the mean SI on the pre-contrast phase. Washout was calculated as (SI_CM − SI_DEL)/SI_CM x 100%, where SI_CM is the mean SI during CM phase and SI_DEL is the mean SI during the late nephrographic phase.

Second-order Haralick Texture Features

Haralick texture features (Table 2) were extracted using an in-house OsiriX (Osirix X, version 5.6, 64 bit, Bernex, Switzerland) plugin that interfaces with pyOsiriX ³³ and an open-source computer vision library for Python (Mahotas) ³⁴. The initial step in this analysis is the discretization of the ROI images. Signal intensities of individual ROI slices were binned into 8-bit integers from the minimum to maximum signal intensity within the ROI at the given slice. Next, the histograms of the intensity values within the individual ROI slices were equalized. The gray-level co-occurrence matrices (GLCMs - also known as a spatial gray-level dependence matrices) were calculated from the discretized signal intensities at each ROI slice. A pixel offset distance of 1 pixel was used. Four GLCMs were calculated corresponding to the four pair-wise directions for the given pixel offset distance. From the GLCMs, second order statistics were generated following the recommendations of the Imaging Biomarker Standardization Initiative (IBSI) ³⁵. For the DCE images, the second order statistics were averaged across the four pair-wise directions. These averaged second order statistics were then exported for further statistical analysis.

Table 2:

Imaging Features

Category	Parameter Type	Abbreviation
First-order Intensity histogram
	Mean
	Standard deviation
	Percentile
	Skewness
	Kurtosis
GLCM based texture (Haralick features)
	Angular second moment	f1
	Contrast	f2
	Correlation	f3
	Sum of squares	f4
	Inverse difference moment	f5
	Sum average	f6
	Sum variance	f7
	Sum entropy	f8
	Entropy	f9
	Difference variance	f10
	Difference entropy	f11
	Information measures of correlation 1	f12
	Information measures of correlation 2	f13

Reference standard

Histopathologic results after nephrectomy (radical or partial) served as the reference standard in all patients. Histopathology slides were reviewed by an experienced uropathologist (PK, with more than 12 years of experience). All tumors were grouped according to the International Society of Urological Pathology grade (ISUP) as high grade (HG, ISUP grade 3-4); or low grade (LG, ISUP grade 1-2) and assigned a pathologic stage. The presence of necrosis was tabulated.

Statistical methods

A total of 85 features were included in the analysis: tumor length, 13 T2 Haralick texture features, 18 DCE 1^st order features (mean, standard deviation, 75^th percentile, 90^th percentile, kurtosis, skewness on each of the 3 normalized DCE phases), DCE wash-out and 52 DCE Haralick texture features (13 features on the PRE and each of the 3 normalized DCE phases). First, a univariable logistic regression was performed to correlate each feature with tumor grade (i.e., HG vs. LG), necrosis (present/absent) and pathologic stage (≤pT2 vs >pT2). Heatmaps were constructed where MRI features and tumors were clustered with complete linkage under Euclidean distance and sorted such that distance between adjacent patients were minimized ³⁶.

Next, a clustering algorithm (VARCLUS procedure, SAS 9.4, SAS Institute, Inc, Cary, NC) was applied to reduce the effect of collinearity (Supplementary Statistical Methods) ^{37, 38}. The variable clustering algorithm was applied to the T2 Haralick texture features, DCE MRI 1^st order features and DCE MRI Haralick texture features, respectively, to maintain interpretability of the clustered variables. Tumor length, wash-out and all cluster features were evaluated univariably by receiver operating characteristic (ROC) curves. Hypothesis tests were carried out to test whether each area under the curve (AUC) was significantly greater than 0.5. The false discovery rate (FDR) adjusted p-values were also reported ³⁹.

Several logistic regression models were proposed to assess the predictability of HG vs LG, necrosis vs no necrosis, and ≤pT2 vs >pT2 stage ccRCC. Stepwise features selection algorithms were used to select from all the cluster features. To avoid overfitting, the criterion for feature selection was cross-validated AUC (via leave-one-out (LOO))⁴⁰, where features enter and leave the model based on the increase or decrease of LOO-AUC. Since tumor size remains an important variable in clinical practice, we assessed three models: 1) a univariable logistic regression model with tumor length only; 2) a multivariable logistic regression model with stepwise selection restricted to include tumor length; 3) a multivariable logistic regression model with stepwise selection with no restriction. All analyses were done in SAS 9.4 (SAS Institute, Inc, Cary, NC).

Results

A total of 102 patients underwent MRI before nephrectomy between August 2012 and August 2017. Twenty five patients were excluded due to non-clear cell RCC histology and 2 due to no completion of the MRI. Of the remaining 75 patients with ccRCC, patients were excluded due to inability to complete the DCE MRI part of study (n=1), had a predominantly cystic tumor (n=1), or for technical reasons (n=3). Therefore, a total of 70 ccRCC (18 female; 52 male; mean age 57.8 ± 10.8 years; range 32-82 years) were included (Figure 1, Table 3). The mean time interval between mpMRI and surgery was 5 days (range, 1–19 days).

Figure 1:

Study flowchart

Table 3:

Patient characteristics

Characteristics (n = 70)	n (%)
Sex
Male	52 (74.3 %)
Female	18 (25.7 %)
Mean age (years)	57.8 ± 10.8; range 32-82
Pathological stage
T1a	26 (37.1 %)
T1b	22 (31.4 %)
T2	0
T3a	17 (24.3 %)
T3b	1 (1.4%)
T3c	1 (1.4 %)
T4	3 (4.3 %)
ISUP Grade
1	0
2	38 (54.3 %)
3	28 (40 %)
4	4 (5.7 %)
Necrosis
Yes	20 (28.6%)
No	50 (71.4%)

ISUP, international society of urological pathology grade; n, number of patients.

Figure 2 illustrates the workflow of data acquisition, extraction, and analysis. The clusters generated by the variable clustering algorithm for T2W and DCE MRI features are detailed in supplemental Table 1. Heat map generated using distance clustering method showing the association between radiomic features and ISUP tumor grade is shown in Figure 3A. A visually prominent representation of clusters of DCE texture parameters were noted and these tend to cluster in patients with either low or high grade tumors (Figure 3B). The diagnostic performance of various histogram and texture features are summarized in Table 4.

Figure 2: Radiomics platform: Workflow for data acquisition, data extraction, and analysis.

Data acquisition: two-dimensional (2D) T2-weighted (T2W) and three dimensional (3D) dynamic contrast enhanced (DCE) magnetic resonance imaging (MRI) datasets were used.

Segmentation: Manual segmentation of the tumor was performed by drawing a region of interest (ROI) in every T2W image displaying the tumor and in a center slice of the DCE acquisition at 4 different time points (i.e., pre-contrast, corticomedullary phase, early nephrographic phase, and late nephrographic phase). Feature extraction: quantitative features were calculated including histogram-based first order statistics, tumor length, and second order statistics; the latter after conversion of each ROI data into a gray level co-occurrence matrix (GLCM). Correlation and data analysis: final histopathology was used as reference standard. Clustering and logistic regression analysis were carried out to test various predictive models of high grade histology.

Figure 3: Correlation of radiomic features and histopathology grades.

Heat map generated using distance clustering method showing the association of all T2W- and DCE-based features evaluated in the study (A) and DCE-MRI texture features (B) with histopathologic grades. Columns represent individual patients and rows represent the histogram and texture imaging features. A bar at the top represents the histopathology grade for each tumor based on the International Society of Urological Pathology (ISUP) grading system (Grey= ISUP grade 1-2 [low grade]; Black= grade 3-4 [high grade]). Several DCE-derived texture features exhibit increased or decreased values in high grade clear cell carcinoma.

Table 4:

Performance of various features

Variable	AUC	95% CI	p-value	FDR
DCETextureClus3	0.75	0.63-0.87	<0.0001	0.0009
DCETextureClus1	0.72	0.59-0.84	0.0006	0.004
Length	0.7	0.58-0.82	0.002	0.008
T2TextureClus3	0.69	0.56-0.81	0.004	0.01
DCEhistClus1	0.68	0.55-0.81	0.005	0.02
DCETextureClus6	0.64	0.51-0.77	0.03	0.08
DCEhistClus4	0.63	0.50-0.77	0.05	0.09
DCETextureClus2	0.63	0.49-0.76	0.06	0.11
T2TextureClus1	0.63	0.49-0.76	0.07	0.11
DCETextureClus5	0.61	0.47-0.74	0.12	0.17
T2TextureClus2	0.59	0.46-0.73	0.18	0.25
DCEhistClus2	0.54	0.40-0.68	0.55	0.69
Washout	0.53	0.39-0.67	0.69	0.80
DCETextureClus4	0.50	0.36-0.64	0.96	0.96
DCEhistClus3	0.49	0.35-0.63	0.92	0.96

AUC, area under the curve; FDR, false discovery rate; CI, confidence interval.

Severe collinearity was observed among multiple imaging features (Figure 4) illustrating the need for clustering. AUC for differentiating LG and HG ccRCC with DCE texture clusters 3 and 1 were 0.75 (95% confidence interval (CI), 0.63-0.87, FDR p- value = 0.0009) and 0.72 (95% CI, 0.59-0.84, FDR p- value = 0.004), respectively. AUC for differentiating LG and HG ccRCC with tumor length was 0.70 (95% CI, 0.58-0.82, FDR p- value = 0.008) (Figure 5A).

Figure 4: Collinearity of imaging features.

Heatmap illustrating the collinearity of DCE MRI histogram parameters, b) DCE MRI texture parameters, and c) T2W texture parameters quantified by Pearson correlation coefficients. Red boxes show positive correlation (>0.7) and blue boxed show the negative correlation (<−0.7).

Figure 5: Diagnostic performance of length and individual clusters of imaging features.

Area under the receiver operating characteristic (ROC) curve and 95% confidence interval of the various cluster components of all features, length and wash-out parameters for the prediction of tumor grade (A), necrosis (B), and >pT2 stage (C).

AUC for differentiating necrosis from no necrosis with DCE histogram clusters 1 and 3 were 0.79 (95% confidence interval (CI), 0.67-0.91, FDR p- value = 0.00002) and 0.73 (95% CI, 0.60-0.86, FDR p- value = 0.003), respectively. AUC for differentiating necrosis from no necrosis with tumor length was 0.71 (95% CI, 0.55-0.86, FDR p- value = 0.04) (Figure 5B).

AUC for differentiating ≤pT2 vs >pT2 stage with DCE texture clusters 3 and 1 were 0.86 (95% confidence interval (CI), 0.75-0.97, FDR p- value < 0.00001) and 0.81 (95% CI, 0.70-0.93, FDR p- value < 0.00001), respectively. AUC for differentiating ≤pT2 vs >pT2 stage with tumor length was 0.81 (95% CI, 0.69-0.94, FDR p- value < 0.00001) (Figure 5C).

Based on the multivariable analysis, three models were constructed and cross-validated to reduce the likelihood of over-fitting ⁴⁰ for the prediction of tumor grade: 1) model 1 included tumor length only; 2) model 2 was constrained to included length and returned with DCE histogram cluster 4 and DCE texture cluster 3; and 3) model 3, an unconstrained model, included DCE histogram cluster 1 and DCE texture cluster 3. The AUC for predicting HG ccRCC of models 1, 2, and 3 were of 0.67 (95% CI 0.54, 0.79), 0.82 (95% CI 0.71, 0.92), and 0.81 (95% CI 0.70 0.91), respectively (Figure 6A). Differences in AUC of model 1 vs. 2 and model 1 vs. 3 were statistically significant (p = 0.02 and p = 0.03, respectively). However, differences in AUC between model 2 and model 3 were not statistically significant (p = 0.82).

Figure 6: Diagnostic performance of predictive models.

(A) Receiver operating characteristic (ROC) curves for prediction of high grade ccRCC with tumor length, a constrained model including length plus DCE MRI clusters 3 and 4, and an unconstrained model with DCE MRI cluster 1 and 3. (B) ROC curves for prediction of necrosis with tumor length, a constrained model including length plus DCE MRI histogram cluster 1, DCE MRI texture cluster 4, and T2W texture cluster 1, and an unconstrained model including DCE MRI histogram cluster 1, DCE MRI texture cluster 1, and DCE histogram cluster 3. The constrained model was superior to length for prediction of necrosis, with a trend toward statistical significance (p=0.04). (C) ROC curves for prediction of >pT2 stage with tumor length, a constrained model including length plus DCE MRI histogram cluster 1, DCE MRI texture cluster 4, and T2W texture cluster 1, and an unconstrained model including DCE MRI histogram cluster 1, DCE MRI texture cluster 1, and DCE histogram cluster 3. No statistical significance was seen between the 3 models.

A constrained model including length plus DCE histogram clusters 1, DCE MRI texture cluster 4, and T2W texture cluster 1 (AUC 0.82, 95% CI 0.72, 0.93) was superior to length (AUC 0.65, 95% CI 0.49, 0.82) in predicting tumor necrosis with a trend towards statistical significance (p=0.04)(Figure 6B). An unconstrained model was not statistical superior for prediction of necrosis. Neither a constrained nor an unconstrained model incorporating histogram or texture features was statistically superior to length at predicting >pT2 stage (Figure 6C).

Discussion

A more general implementation of AS in patients with renal masses is impeded by the lack of reliable predictors of disease aggressiveness, particularly in those patients with larger tumors. Clear cell neoplasms are more likely to metastasize than other histologic subtypes of RCC ^{8, 41}. A correlation exists between histologic grade in RCC and increasing risk of metastasis, and the prevalence of high grade histology increases with tumor size ^{5, 42}. A definitive diagnosis of ccRCC can be achieved with a percutaneous biopsy with high level of accuracy, however the prediction of tumor grade is unreliable ¹⁴. A recent meta-analysis of percutaneous renal mass biopsies revealed a negative predictive value of 63.3% and 16% upgrade rate of masses with low-grade histology at biopsy ¹⁴. Similarly, the presence of tumor necrosis, an independent prognostic variable ¹⁵, can be missed on a biopsy. This is not surprising given that tumor biopsies only represent a minute portion of the tumor and intratumoral heterogeneity is a hallmark of ccRCCs ^{12, 13}. Importantly, some of the most commonly used prognostication models including tumor grade and necrosis such as SSIGN model require prior surgery ^{43, 44}. The evaluation of renal tumors with cross-sectional imaging offers an opportunity to assess the entire tumor or, at least, a larger area of the tumor. Our analysis assessing tumor heterogeneity in the entire tumor on T2W imaging or on a single, center slice on contrast-enhanced images offers new evidence toward this goal. Our predictive model using a number of texture features on MRI had a reasonable diagnostic performance for predicting ccRCC (AUC of 0.8), which was superior to tumor maximum dimension, a common measure used in clinical practice.

To our knowledge, this is the first report using radiomics on mpMRI data to predict high grade histology in ccRCC. Chandarana et al. ²³ reported a semi-automated segmentation approach and histogram analysis of the whole lesion enhancement to differentiate RCC subtypes with promising results. Our study differs in that we included both histogram analysis and texture features and that we focused instead on the prediction of high grade histology in ccRCC. It could be argued that our results would only be applicable to patients with known ccRCC, such as those who have had a previous biopsy. For these patients, a radiomic assessment indicating a high likelihood of high grade histology could help make management decisions if the previous biopsy indicated low grade histology. An incorrect risk stratification due to sampling error would be suspected in such patient, particularly in the presence of a large, heterogeneous tumor.

However, radiomic analysis would also be applicable to patients with suspected ccRCC without histologic confirmation. Canvasser et al ²² reported a sensitivity and specificity of 78% and 80%, respectively, for the prediction of ccRCC using a predefined algorithm for interpretation of MRI examinations of patients with cT1a renal tumors to generate a ccLS. Similarly, Kay et al ⁴⁵ used a similar algorithm to predict histologic subtypes in patients with cT1a renal masses. The diagnostic algorithm is based primarily on signal characteristics of the mass on T2-weighted images and contrast enhanced images during the corticomedullary phase and these two variables have been shown to have the highest inter-observer reproducibility ⁴⁵. However, neither report assessed the ability to diagnose high grade histology. The same MRI variables were used in our proposed radiomics approach as these are standard acquisitions in most multiparametric MRI examinations. Radiomic analyses on these sequences could add to the non-invasive assessment of renal masses previously proposed by Canvasser et al.

The role of diffusion weighted imaging (DWI) in the characterization of renal masses is still under investigation ⁴⁶. Rosenkrantz et al ⁴⁷ evaluated the utility of DWI for prediction of high grade histology in ccRCC. Using high b values of 400 s/mm² and 800 s/mm², they found a statistically significant lower ADC in higher grade tumors compared to lower grade tumors. However, only tumor necrosis and perinephric fat invasion remained statistically significant at multivariable analysis for the prediction of high grade ccRCC. We did not include DWI data in our analysis. Further work is necessary to assess the value of radiomics in DWI for prediction of tumor grade.

We found a better prediction of tumor grade and necrosis with combination of several histogram and texture features of T2W and DCE MRI. If validated in largest series, an accurate determination of tumor grade and/or necrosis prior to surgery can have several important implications. First, it may facilitate early intervention for high-risk smaller tumors that may have otherwise been considered for surveillance. Second, it could enable assessment of the performance of extended lymphadenectomy at time of surgery in higher risk tumors, a surgical intervention that is not widely accepted. Third, it could allow enrollment into neoadjuvant therapy trials minimizing the reliance on biopsy. Four, it could provide a guide for directing biopsies to the most aggressive area of a renal mass for better risk stratification. Lastly, it could assist in the implementation of AS protocols in larger, heterogeneous tumors.

Our study has several limitations. First, a training set and external validation set were not used because the total number of patients in our cohort was relatively small. A leave-one-out crossvalidation method was used to decrease the risk of over-fitting ⁴⁰. Second, we only used Haralick features for texture analysis. The number of reported texture features continues to increase with over 150 features now recommended by the IBSI standards ³⁵ however these were not available to us at the time of completion of this work. Further studies including these recommended texture features are needed. Third, we did not perform a one-to-one correlation between texture features and histopathologic grade in the whole tumor. Such work is labor intensive and requires careful co-registration between the surgical specimen (i.e. preferably with whole mount histology) and in vivo imaging. While this is feasible with the use of patient-specific 3D molds ⁴⁸ it was not done routinely in our cohort. Additional research in this area would be important and could help elucidate the potential role of MRI in directing percutaneous biopsy (i.e., selection of most suspicious area for high grade histology in the mass). Fourth, although the MRI acquisition protocol included a DCE MRI protocol, no quantitative parameters such as K^rans or K_ep were utilized, which have been shown to be helpful in predicting high grade ccRCC in T1b tumors in preliminary studies ⁴⁹. The reliability of such methods with standard acquisition techniques is challenged by the presence of intra-abdominal motion. We chose pre-contrast and 4 phases from the DCE MRI acquisition that match those used in most clinical protocols ⁵⁰ this should facilitate the clinical implementation of our methods. Further research using novel 4D acquisitions for DCE MRI is needed to better understand the role these quantitative pharmacokinetic parameters in the evaluation or renal tumors. Fifth, we used single slice analysis instead of whole tumor volumetric analysis.

Conclusion

In summary, our preliminary radiomic analysis of T2-weighted and dynamic contrast-enhanced MR images was superior to tumor size for the prediction of high grade clear cell renal cell carcinoma in our cohort. Further work using an extended number of texture features and validation of these findings in an independent cohort is necessary.

Clinical Practice Points:

1. Radiomics analysis improved the prediction of high grade histology in clear cell renal cell carcinoma compared to prediction based on tumor size (area under the receiver operating curve [AUC] of 0.82 [95% CI, 0.71-0.92] versus 0.67 [95% CI, 0.54-0.79], respectively) (p=0.02).

2. Radiomics analysis improved the prediction of necrosis in clear cell renal cell carcinoma compared to prediction based on tumor size (AUC of AUC 0.82 [95% CI 0.72, 0.93] versus AUC 0.65 [95% CI 0.49, 0.82], respectively) (p=0.04).

3. Incorporation of radiomics analysis in the assessment of MRI in patients with clear cell renal cell carcinoma has the potential to assist in management decisions of intervention vs. active surveillance, the selection patients for neoadjuvant therapy, and to provide a guidance for the need of percutaneous biopsy in patients with primary ccRCC.

Abbreviations:

RCC

renal cell carcinoma

ccRCC

clear cell renal cell carcinoma

MRI

magnetic resonance imaging

DCE-MRI

dynamic contrast-enhanced MRI

GLCM

gray-level co-occurrence matrix

ISUP

International Society of Urological Pathology

ROI

region of interest

mpMRI

multiparametric magnetic resonance imaging

PRE

pre-contrast

CM

corticomedullary phase

NG

early nephrographic phase

DEL

late nephrographic/delayed phase

FDR

the false discovery rate adjusted p-values

LG

low grade

HG

high grade

pT2

pathological T2 stage

ROC

receiver operating characteristic

AUC

area under the curve

T2W

T2-weighted

CI

confidence interval

LOO

leave-one-out

ccLS

clear cell likelihood score