Genitourinary Imaging Original Research

Transcription

1 Genitourinary Imaging Original Research Lubner et al. CT Texture Features and RCC Genitourinary Imaging Original Research Meghan G. Lubner 1 Nicholas Stabo 1 E. Jason Abel 2 Alejandro Munoz del Rio 1 Perry J. Pickhardt 1 Lubner MG, Stabo N, Abel EJ, Munoz del Rio A, Pickhardt PJ Keywords: CT texture, heterogeneity, renal cell carcinoma DOI: /AJR Received August 20, 2015; accepted after revision January 16, Based on presentations at the Society of Abdominal Radiology 2014 annual meeting, Budapest, Hungary and the American Urologic Association 2015 annual meeting, New Orleans, LA. 1 Department of Radiology, University of Wisconsin School of Medicine and Public Health, E3/311 Clinical Sciences Center, 600 Highland Ave, Madison, WI Address correspondence to M. G. Lubner (mlubner@uwhealth.org). 2 Department of Urology, University of Wisconsin School of Medicine and Public Health, Madison, WI. AJR 2016; 207: X/16/ American Roentgen Ray Society CT Textural Analysis of Large Primary Renal Cell Carcinomas: Pretreatment Tumor Heterogeneity Correlates With Histologic Findings and Clinical Outcomes OBJECTIVE. The purpose of the present study is to determine whether CT texture features of newly diagnosed primary renal cell carcinomas (RCCs) correlate with pathologic features and oncologic outcomes. MATERIALS AND METHODS. CT texture analysis was performed on large (> 7 cm; mean size, 9.9 cm) untreated RCCs in 157 patients (52 women and 105 men; mean age, 60.3 years). Measures of tumor heterogeneity, including entropy, kurtosis, skewness, mean, mean of positive pixels, and SD of pixel distribution histogram were derived from multiphasic CT using various filter settings: unfiltered (spatial scaling factor, 0), fine (spatial scaling factor, 2), medium (spatial scaling factor, 3 4), or coarse (spatial scaling factor, 5 6). Texture values were correlated with histologic subtype, nuclear grade, pathologic stage, and clinical outcome. RESULTS. When a coarse filter setting (spatial scaling factor, 6) was used, entropy on portal venous phase CT images was positively associated with clear cell histologic findings (odds ratio [OR], 134; 95% CI, ; p < 0.001) and was negatively associated with non clear cell subtype findings (papillary spatial scale factor, 6; OR, 0.016; 95% CI, ; p < 0.001). ROC curve analysis for entropy (on portal venous phase images obtained with a spatial scaling factor of 6) revealed an AUC of (95% CI, ) for clear cell histologic findings, with similar values noted for non clear cell histologic findings. The mean of positive pixels and the SD of the pixel distribution histogram were statistically significantly associated with histologic cell type in a similar fashion. Entropy, the SD of the pixel distribution histogram, and the mean of positive pixels were associated with nuclear grade, most prominently when fine or medium texture filters were used (p < 0.05). There was a statistically significant association of texture features noted on unenhanced CT, including the SD of the pixel distribution histogram, the mean of positive pixels, and entropy, with the time to disease recurrence and death due to disease (e.g., for entropy noted on unenhanced CT images obtained with a spatial scaling factor of 6, the hazard ratio was 3.49 [95% CI, ]; p = 0.002). CONCLUSION. CT texture features (in particular, entropy, the mean of positive pixels, and the SD of the pixel distribution histogram) are associated with tumor histologic findings, nuclear grade, and outcome measures. The contrast phase does seem to affect heterogeneity measures. P atients with renal cell carcinoma (RCC) who have large primary tumors are more likely to have metastatic disease at initial presentation and an increased risk of metastases developing after treatment. For most patients with metastatic RCC, treatment does not include nephrectomy [1, 2], and RCC is diagnosed using percutaneous biopsy, which has limited accuracy in identifying adverse pathologic features in large metastatic tumors [3, 4]. This inaccuracy of percutaneous biopsy performed for patients with large metastatic tumors is likely the result of a sampling error of large heterogeneous primary RCC tumors [4]. However, there is a critical need for predictive information about the RCC histologic subtype, tumor grade, and other prognostic features, to help guide treatment decisions for patients with metastatic RCC [5], especially patients with aggressive subtypes [6]. The choices of systemic therapy may vary significantly, depending on whether clear cell or non clear cell RCC is diagnosed [7] or whether patients have sarcomatoid dedifferentiation [8]. Alternative methods must be developed to obtain prognostic information from large heterogeneous RCC tumors and to guide 96 AJR:207, July 2016

2 CT Texture Features and RCC the choice of systemic therapies, because RCC tumors are known to be inherently heterogeneous, identification of features prognostic of RCC tumors has not been reliably shown with the use of biopsy, and metastatic RCC is not treated surgically in most patients. Using imaging features as prognostic biomarkers has several advantages. First, imaging can be performed noninvasively. Second, CT and MRI are already used as part of clinical staging for RCC. Third, a variety of studies have reported that both CT and MRI can identify potentially useful imaging biomarkers, including diffusion restriction, apparent diffusion coefficient values, and features of dynamic contrast enhancement [9 16]. Another potentially useful emerging biomarker is CT tumor texture, which has shown promise in predicting pathologic features, overall survival, and response to A therapy for a variety of tumor types [17 21]. Texture analysis provides an assessment of tumor heterogeneity by analyzing the distribution of and association with pixel or voxel gray levels in the image [22]. Although direct pathologic correlates have not been determined, it has been suggested that texture provides information on the tumor microenvironment [23 25]. There has been some assessment of the usefulness of texture features in the evaluation of metastatic RCC treated with tyrosine kinase inhibitors [26]; however, to our knowledge, there are limited published data that assess the value of texture analysis for untreated primary RCC. Raman et al. [27] showed that texture features can be used to reliably discriminate between different types of renal masses (e.g., clear cell renal carcinoma, papillary renal carcinoma, oncocytoma, and renal cyst). However, it is important to correlate texture findings both with pathologic findings for primary tumors that have been treated surgically and with oncologic outcomes before this method, for prognostic purposes, is applied to the majority of patients with metastatic RCC who are treated without nephrectomy. Therefore, the purpose of the present study is to determine whether the CT texture features of large primary RCCs are associated with pathologic tumor features and oncologic outcomes. Materials and Methods The present study was approved by the institutional review board at the University of Wiscon- D E F Fig year-old man with large renal cell carcinoma (RCC) on right side. A and B, Portal venous phase contrast-enhanced transverse (A) and coronal (B) CT images show tumor (arrows). C, Texture analysis software image shows ROI (blue line) outlining tumor. D F, Color texture overlays of tumor outlined by ROI (blue line) show images with fine filtering (spatial scaling factor, 2) (D), medium filtering (spatial scaling factor, 3 4) (E), and coarse filtering (spatial scaling factor, 5 6) (F). Quantitative output occurs on basis of this filtration and spatial scaling. Texture parameters of this patient would have correctly identified tumor as clear cell subtype (for portal venous phase images obtained with spatial scaling factor of 6, mean of positive pixels was 26.9, SD of pixel distribution histogram was 49.1, and entropy was 5.09; Table 2) with nuclear grade of 3 (for portal venous images obtained with spatial scaling factor of 2, mean of positive pixels was 37.2 and entropy was 5.2, whereas spatial scaling factor of 3 resulted in mean of positive pixels of 33.67; Table 3), as confirmed by pathologic findings after surgery. B C AJR:207, July

3 Lubner et al Entropy Sensitivity (%) Non Clear Cell Subtype AUC = (95% CI, ) Specificity (%) sin School of Medicine and Public Health and was HIPAA compliant. The need for informed consent was waived. Clear Cell Subtype C Density Patient Population The CT images of 157 patients (52 women and 105 men; mean age, 60.2 years) with large (> 7 cm) RCCs were obtained from the surgical database of the department of urology and were retrospectively reviewed. All patients in the cohort had a CT scan performed before undergoing surgery or receiving any other treatment. Subsequent surgical removal of the primary tumor and pathologic analysis that included histologic subtyping and nuclear grading were performed for all patients. The dates of imaging and surgery were recorded. Data from clinical and imaging follow-up examinations were also obtained for all patients from the PACS and the electronic medical records. The date of death was also confirmed on the basis of data in public records. Overall survival was calculated as the time from the date of surgery to either the date of the last follow-up examination or the date of death. Patient status (i.e., no evidence of disease, alive with disease, deceased due to disease, or deceased due to other causes) at the time of the last clinical or imaging follow-up examination was also recorded. CT Images and Analysis The CT images obtained on the date closest to the date of surgery were selected for analysis. Examinations were classified into series on the basis of the contrast phase (unenhanced [n = 107], arterial phase [n = 29], portal venous or nephrographic phase [n = 135], and delayed phase [n = 99]). Measurements were performed separately for each phase, and CT texture findings were recorded according to contrast phase. A total of 125 of 157 patients (80%) had images obtained in more than one contrast phase, but very few (5/157; 3%) had images obtained in all four phases. A total of 92 of 157 patients (58%) had both unenhanced phase and either portal venous or nephrographic phase images obtained. Comparisons were performed for images from the same contrast phase only. Given that the two Entropy Non clear cell subtype Clear cell subtype Not chromophobe cell type Chromophobe cell type Not papillary cell NOS Papillary cell NOS A B Fig. 2 Correlations between histologic subtypes and entropy. A, Box-whisker plot shows differentiation of clear cell histologic subtype (blue circles) from non clear cell histologic subtype (red circles), according to entropy noted on portal venous or nephrographic phase images obtained with use of coarse filter settings (spatial scaling factor, 6; p < 0.001). Horizontal lines within boxes denote mean values, and vertical lines and whiskers denote 95% CIs. B, Density plot shows findings similar to those noted for entropy on portal venous phase images obtained with spatial scaling factor of 6. Findings clearly separate clear cell histologic subtype (dark blue line) from non clear cell histologic subtype (light blue line). Note that total of values associated with chromophobe cell type (dark green line) and papillary cell type (red line) equals non clear cell values (light blue line). NOS = not otherwise specified. C, ROC analysis shows strong correlation between clear cell histologic subtype and entropy noted on portal venous phase images obtained with coarse filter settings (spatial scaling factor, 6) on portal venous phase image, with AUC of (95% CI, ) noted. Similar AUC values were seen for mean of positive pixels and SD of pixel distribution histogram on portal venous phase image obtained with coarse filter settings. Black line is ROC curve, and gray line is reference. largest series of images were obtained in the unenhanced images and portal venous or nephrographic phase images, images obtained in these phases were the focus of the analysis. The CT technique was slightly heterogeneous, with 92 of 157 CT examinations performed at institutions other than the study institution. All CT scans were performed using MDCT scanners, and the imaging parameters were as follows: a tube potential of kv (with 122 of the 157 scans using a tube potential of 120 kv) and a matrix of Most CT scans were performed using automated or variable tube potential, and the slice thickness used for 125 of 156 scans was 2 5 mm. There are data to suggest that the method of texture measurement used in the present study is resistant to differences in technique [28]. After review of the CT images on a PACS, a single slice at the level of the overall largest transverse diameter of the tumor was selected. If a patient had images obtained in more than one contrast enhancement phase, the same slice level was selected for all phases. All images were re- 98 AJR:207, July 2016

4 CT Texture Features and RCC viewed, and slices were selected by experienced attending abdominal imaging radiologists, one of whom had 18 years of experience and one of whom had 10 years of experience. Single-slice images were then sent to a commercially available texture analysis software program (TexRAD, TexRAD). There are data to suggest that a single slice is sufficient to conduct analysis with this software [29, 30]. With the use of the software, an ROI was manually drawn around the outer margin of the tumor by a single reader, a medical student who was under the supervision of the two abdominal imaging radiologists. The software uses an initial filtration step and a Laplacian of gaussian spatial bandpass filter to selectively extract features of different sizes and variations in intensity [23, 26]. This allows features to be seen on a series of images derived using texture filter settings ranging from fine (spatial scaling factor, 2; width, 4 pixels; object radius, 2 mm) to coarse (spatial scaling A C factor, 5 6; width, 12 pixels; object radius, 6 mm) [25, 31] (Fig. 1). The software output includes information on a variety of histogram characteristics, including mean pixel attenuation, the SD of the pixel distribution histogram, entropy, the mean of positive pixels, the skewness (i.e., asymmetry) of the pixel histogram, kurtosis (i.e., peakness) of the pixel histogram, and the percentage of positive pixels noted in association with the use of each spatial scaling factor. These values were recorded for each study on the basis of the contrast phase (with all data from each contrast phase evaluated separately) and were subsequently analyzed. Texture features were correlated with histologic subtype, nuclear grade, presence of sarcomatoid features, presence of metastatic disease, time to disease recurrence, and time to death from disease. Note that nuclear grade was not applied to chromophobe RCCs, so such RCCs are categorized as having an unknown grade. B Fig year-old woman with renal cell carcinoma (RCC) on right side. A C, Unenhanced (A), portal venous phase (B), and delayed phase (C) CT images show right upper pole RCC (arrows). Texture features corroborate non clear cell or chromophobe histologic subtype (for portal venous phase image obtained with spatial scaling factor of 6, mean of positive pixels was 16.2, SD of pixel distribution histogram was 20.15, and entropy was 4.41; Table 2). Statistical Analysis Continuous variables were summarized using descriptive statistics. Frequencies and corresponding percentages were determined for categoric variables. To assess the association between CT texture features and clinical data or outcomes, different methods were used, depending on the type of response being considered. For association of texture parameters with numeric and ordinal variables (values such as nuclear grade), ordinary least squares or simple linear regression analysis was used. Analysis was performed for each of the six parameters at each filter level and for each contrast phase (i.e., the unenhanced phase, arterial phase, portal venous or nephrographic phase, and delayed phase). Binary variables (such as RCC subtype, presence of sarcomatoid features, and presence of metastatic disease) were analyzed using logistic regression. Odds ratios (ORs) were reported with 95% CIs. ROC curves and 95% CIs, calculated on the basis of the DeLong method, were also obtained. AJR:207, July

5 Lubner et al. For survival or time-to-event responses, Cox proportional hazards regressions were fitted; hazard ratios (HRs) and 95% CIs were also obtained. To gain additional insight, predictors were grouped into tertiles, and the corresponding Kaplan-Meier survival estimates were obtained and plotted for each group. Given that six main texture parameters were included, a Bonferroni correction for multiple testing was performed by dividing 0.05 by 6, with p = denoting statistical significance. Although each texture parameter was evaluated at multiple spatial scaling factors, these are not thought to be independent, and we therefore did not account for them in our multiple testing correction [29]. Multiple masses from the same patient were considered independently. TABLE 1: Demographic and Clinical Characteristics of Patients with Renal Cell Carcinoma Characteristic All Patients (n = 157) Patients With Portal Venous or Nephrographic Phase Images (n = 135) All statistical graphics and computations were obtained using R version (R Foundation). Results The demographic and clinical characteristics of the patient cohort are summarized in Table 1. All tumors that were analyzed were large (> 7 cm; mean [± SD] size, 9.9 ± Patients With Unenhanced Images (n = 109) Patients With Both Portal Venous or Nephrographic Phase and Unenhanced Phase Images (n = 92) Age (y), mean Sex, no. of patients Female Male Tumor diameter, cm Mean ± SD 9.9 ± ± ± ± 2.8 Median Range 7 22 Tumor histologic subtype Clear cell Papillary Chromophobe Unclassified Nuclear grade Unknown Sarcomatoid features, no. (%) of patients 13 (8) 12 (9) 8 (7) 7 (8) Metastatic disease present at surgery, no. (%) of patients 43 (27) 36 (27) 28 (26) 23 (25) Clinical follow-up duration, mo Mean ± SD 28.1 ± ± ± ± 33 Median Range Patient status Deceased Renal cell carcinoma related Unknown or alternate cause Alive With no evidence of disease With disease Overall survival, mo Mean ± SD 31.5 ± ± ± ± 32.7 Median Range AJR:207, July 2016

6 CT Texture Features and RCC Mean of Positive Pixels SD Entropy A Fig. 4 Associations among nuclear grade of renal cell carcinoma and CT texture features. B C A C, Regression plots show negative association between nuclear grade of renal cell carcinoma and mean of positive pixels on portal venous phase image obtained with use of fine texture filter setting (spatial scaling factor, 2; p = 0.002) (A), SD of pixel distribution histogram on portal venous phase image obtained with use of fine texture filter setting (spatial scaling factor, 2; p = ) (B), and entropy on unenhanced CT image obtained with use of medium texture filter setting (spatial scaling factor, 3; p = 0.005) (C). Circles are data points, and black line is line of best fit. 2.9 cm). Most of the tumors were of the clear cell histologic subtype (n = 131), with smaller numbers of the papillary and chromophobe subtypes noted. A spectrum of nuclear grades was noted for the tumors (papillary and clear cell subtypes), with the majority of tumors assigned a nuclear grade of 2 or higher (Table 1). Nuclear grade was not assigned to chromophobe tumors. Thirteen of 157 patients (8%) had sarcomatoid features. Fortythree patients (27%) had metastatic disease at the time of surgery, which was therefore performed for cytoreductive purposes. The mean time between the evaluation of CT images and the date of surgery was 1.0 month (SD, 0.7 month; median, 0.9 month; range, months). The mean duration of clinical follow-up after surgery was 28.1 months (range, months). At the date of the last follow-up, 56 patients (36%) were dead and 101 were alive. Of the deceased patients, 34 (61%) died of a cause related to their RCC, whereas the other 22 died of unknown or alternate causes. Of those patients who were alive at the last follow-up, 82 (81%) had no evidence of disease and 19 were alive with disease. The mean overall survival was 31.5 months (median, 20 months). The CT texture features noted on portal venous or nephrographic phase images showed the strongest association with histologic findings. Entropy was positively associated with the clear cell histologic subtype and was most often seen when the coarse texture filter setting (spatial scaling factor, 6; OR, 134; 95% CI, ; p < 0.001) was used. However, this effect persisted at all filter texture settings. Entropy was negatively associated with the papillary cell type, most prominently in association with the coarse texture filter setting (spatial scaling factor, 6; OR, 0.016; 95% CI, ; p < 0.001), but persisted across all filter settings. A similar, albeit not quite statistically significant, negative association between entropy and the chromophobe cell type was seen when the medium and coarse filter settings (spatial scaling factor, 3; OR, 0.004; 95% CI, to 0.35; p = 0.015) were used, with the effect also persisting when the fine filter setting was used. When a spatial scaling factor of 6 was used to obtain portal venous phase images, ROC curve analysis of entropy revealed an AUC of (95% CI, ) for the clear cell histologic subtype, with similar values noted for the chromophobe and papillary histologic subtypes (Figs. 1 and 2). With regard to entropy noted on portal venous or nephrographic phase images obtained with a spatial scaling factor of 6 (coarse features), the clear cell histologic subtype had a mean entropy value of 5.3 ± 0.4 and a median value of 5.3, compared with the non clear cell histologic subtype, which had a mean value of 4.5 ± 0.3 and a median value of 4.4. With the use of a threshold of 4.86 for entropy, when portal venous or nephrographic phase images were obtained at a spatial scaling factor of 6 (coarse features), clear cell histologic subtypes could be distinguished from non clear cell histologic subtypes with a sensitivity of 93% and a specificity of 85% (Figs. 1 and 2; Table 2). Additional data on median and mean values for other promising texture parameters and filters, as well as data on thresholds and sensitivity and specificity pairs from ROC analysis, are presented in Table 2. On portal venous or nephrographic phase images, other texture features, including the SD of the pixel distribution histogram and TABLE 2: Association Between Texture Features and Tumor Histologic Subtypes Noted on Portal Venous or Nephrographic Images Obtained With a Coarse Filter Setting (Spatial Scaling Factor, 6), with Sensitivity and Specificity Pairs for ROC Curves Texture Parameter and Histologic Subtype Mean ± SD Median Value Threshold Sensitivity (%) Specificity (%) Mean of positive pixels Clear cell 47.0 ± Non clear cell 19.0 ± SD Clear cell 59.8 ± Non clear cell 25.6 ± Entropy Clear cell 5.3 ± Non clear cell 4.5 ± Note SD = SD of the pixel distribution histogram. AJR:207, July

7 Lubner et al. TABLE 3: Mean (± SD) CT Texture Values, as Stratified by Contrast Phase Texture Parameter the mean of positive pixels, showed a positive association with the clear cell histologic subtype, particularly when coarse filter texture settings were used, but, again, persisting across all filter settings; for example, when such images were obtained with a spatial scaling factor of 6, the SD of the pixel distribution histogram had an OR of 906 (95% CI, 40 20,340; p < 0.001). The negative association with papillary and chromophobe cell types was also seen for the SD of the pixel distribution histogram and the mean of positive pixels; for example, when the papillary type was noted on a portal venous or nephrographic phase image obtained with a spatial scaling factor of 6, the SD of the pixel distribution histogram had an OR of 0.89 (95% CI, ; p < 0.001) (Fig. 3). A weak negative association for which statistical significance was not noted was seen between the percentage of positive pixels and the presence of sarcomatoid features on unenhanced and delayed contrast-enhanced CT Texture Filter (Spatial Scaling Factor) 1 images obtained with the use of medium-tocoarse filter texture settings (for percentage of positive pixels, when delayed phase images were obtained with a spatial scaling factor of 4, OR was [95% CI, to 0.01]; p = 0.01). There was a negative association between similar texture features (entropy, the mean of positive pixels, and the SD of the pixel distribution histogram) on images from multiple contrast phases (i.e., portal venous or nephrographic phase, unenhanced, and delayed phase images) and nuclear grade, when nearly all texture filter settings were used; however, the association was most prominent when the fine and medium texture filter settings were used (Table 3 and Figs. 1, 3, and 4). Nevertheless, this finding did not reach statistical significance when corrected for multiple testing. The mean values for nuclear grades 1 4 were nonoverlapping for similar texture features (Table 3). There was a weak association between the presence of metastatic disease at the time of p Portal venous or nephrogenic Mean of positive pixels Fine (2) 44.2 ± ± ± ± Portal venous or nephrogenic Mean of positive pixels Medium (3) 38.9 ± ± ± ± Portal venous or nephrogenic Entropy Fine (2) 5.38 ± ± ± ± Unenhanced Entropy Fine (2) 5.17 ± ± ± ± TABLE 4: Correlation Between Texture Features Seen on Unenhanced CT Images and Survival of Patients With Renal Cell Carcinoma Texture Parameter Texture Filter (Spatial Scaling Factor) Outcome Hazard Ratio 95% CI p SD Coarse (6) Death due to disease SD Coarse (5) Death due to disease Mean of positive pixels Coarse (6) Death due to disease SD Medium (4) Death due to disease Mean of positive pixels Coarse (5) Death due to disease Mean of positive pixels Medium (4) Death due to disease Mean of positive pixels Medium (3) Death due to disease SD Medium (3) Death due to disease Entropy Coarse (6) Death due to disease Entropy Coarse (6) Time to recurrence SD Coarse (6) Death from all causes Entropy Coarse (5) Time to recurrence Mean of positive pixels Unfiltered (0) Death from other causes SD Unfiltered (0) Death due to disease Note SD = SD of the pixel distribution histogram. surgery and the observation of entropy on both unenhanced (spatial scaling factor, 5; HR, 4.8; 95% CI, ; p = 0.02) and delayed contrast-enhanced (spatial scaling factor, 6; HR, 3.3; 95% CI, ; p = 0.03) CT images obtained with coarse filter settings. There was also a weak association between the SD of the pixel distribution histogram and the presence of metastatic disease when coarse filter settings were used for unenhanced images (spatial scaling factor, 6; HR, 1.1; 95% CI, ; p = 0.03). The strongest association between texture features and survival measures was seen for the unenhanced images, primarily with the same three texture features (the SD of the pixel distribution histogram, entropy, and the mean of positive pixels) previously mentioned. This was again seen at almost all filter settings, but it was most prominently associated with coarse and medium filter features (Table 4 and Fig. 5). The mean value of the pixel histogram was also found to be correlated with death due to all causes and with time to disease recurrence when fine filter texture settings were used. When predictors were grouped into tertiles, Kaplan-Meier plots showed separation of the three groups on the basis of death due to disease and time to disease recurrence (Fig. 5), with shorter survival expected as the value of the mean of positive pixels was increased. Discussion Large RCC tumors are prone to genetic, epigenetic, and pathologic heterogeneity [32], which makes sampling errors more common and also decreases the accuracy of renal mass biopsy in patients with large RCC tumors [3, 4]. CT and MRI are commonly used in the staging of RCC, before treatment options are considered. A variety of promising imaging features have been reported in association with contrast-enhanced CT and MRI, but few studies have investigated the use of texture analysis for patients with large RCC tumors before surgery. Raman et al. [27] found that CT texture analy- 102 AJR:207, July 2016

8 CT Texture Features and RCC sis, when used in conjunction with random forest modeling, shows promise as a tool for discriminating benign from malignant renal lesions as well as some types of malignant lesions (clear cell vs papillary RCC). Our results agree with these findings and then go a step further by suggesting that CT texture analysis in particular, analysis of texture features such as entropy, the SD of the pixel distribution histogram, and the mean of positive pixels may be useful for characterizing histologic cell type, determining nuclear grade, and predicting survival for patients with large RCCs. These techniques may be used to guide clinical decision making or to improve the targeting of biopsies for large heterogeneous tumors. For multiple tumor types, an association has been shown between CT texture features and the pathologic subtypes or even genetic profiles (e.g., KRAS viral oncogene and epidermal growth factor receptor) of the tumor [18, 33, 34]. In addition, CT texture features repeatedly have been shown to correlate with Survival Rate (%) Months clinical outcomes, such as response to therapy and survival [17, 19 21, 35 38]. For RCC in particular, Goh et al. [26] showed that changes in tumor texture in patients with metastatic RCC treated with tyrosine kinase inhibitors were a better predictor of progression-free survival than more established response assessment measures (e.g., Response Evaluation Criteria in Solid Tumors, Choi criteria, and modified Choi criteria). However, there is a relative paucity of data examining the value of CT texture features in assessing primary untreated RCCs at the time of initial diagnosis. Our data show that even before treatment, the texture features of large RCCs, particularly when considered in combination with other imaging and biopsy data, may aid in improving tissue characterization and, potentially, in identifying high-risk tumors and patients. This kind of pretreatment information is easy to obtain and could be valuable in patient counseling and treatment planning. Texture analysis extracts from CT images spatial information that may not be readily ( ) ( ) ( ) A perceptible to the radiologist s eye. Although it is not totally clear what texture represents in the tumor microenvironment, some data suggest that features such as tumor metabolism, hypoxia, and angiogenesis may be associated with tumor heterogeneity [22 25]. For example, observation of tumor heterogeneity at medium and coarse texture filter settings correlated with hypoxia and angiogenesis in primary non small cell lung cancer and colorectal cancer [24, 39]. Simulated models have shown that vascular heterogeneity may be reflected in CT measurements of entropy in the liver [40]. This type of information might be particularly valuable with regard to RCC. A variety of studies have shown that different RCC subtypes have different levels of vascularity and microvessel density reflected by different enhancement patterns [14, 15, 41, 42]. At the genomic level, clear cell RCC has long been associated with mutation of the von Hippel Lindau tumor suppressor gene, which can lead to upregulation of the hypox- Fig. 5 Correlation between CT texture features and survival outcome. A, Kaplan-Meier survival plot shows separation of tertiles according to mean of positive pixels on unenhanced CT images obtained with unfiltered filter setting (spatial scaling factor, 0) plotted against time to death from disease (p = 0.008). For tertile with longest survival, mean of positive pixels ranges from 11.1 to 31.3 on unenhanced CT images obtained with spatial scaling factor of 0. Key in upper right corner of survival plot shows ranges for all tertiles. B and C, 49-year-old man with grade 2 clear cell renal cell carcinoma (RCC) on right side who survived for 99 months. Unenhanced (B) and portal venous phase (C) CT images show RCC (arrows). For this patient, texture features supported pathologic classifications. For clear cell RCC on portal venous phase image obtained with spatial scaling factor of 6, mean of positive pixels was 54.7, SD of pixel distribution histogram was 84.02, and entropy was 5.72 (Table 2). For nuclear grade of RCC on portal venous phase image obtained with spatial scaling factor of 2, mean of positive pixels was 44.2 and entropy was 5.37, whereas, for nuclear grade of RCC on portal venous phase image obtained with spatial scaling factor of 3, mean of positive pixels was 50.9 (Table 3). In addition, patients who, on unenhanced CT images obtained with a spatial scaling factor of 0, had mean of positive pixels of 29.9, were included in first tertile, which was associated with best survival outcome in Kaplan-Meier survival plot. B C AJR:207, July

9 Lubner et al. ia-inducible factors that can trigger angiogenesis. Similarly, recent advances in wholegenome sequencing of clear cell RCC have shown that there are a variety of additional mutations that may be associated with advanced stage, advanced grade, and, possibly, worsened cancer-specific survival [13, 43, 44]. The presence of sarcomatoid features reflects extremely high-grade malignancy and is associated with a markedly worse prognosis [6, 45]. These types of tumoral changes and mutations may contribute to some of the differences captured in the texture measures of RCC seen in the present study. In terms of histologic analysis, entropy and the SD of the pixel histogram were positively correlated with the clear cell histologic subtype and were negatively correlated with papillary and chromophobe cell types, particularly when coarse spatial scaling filters were used for contrast-enhanced imaging. These findings suggest that larger more macroscopic areas of heterogeneity may be seen more commonly in clear cell types. In addition, the mean of positive pixels, particularly on portal venous phase images, also was positively correlated with the clear cell histologic subtype, possibly in association with areas of increased enhancement or hemorrhage. However, when finer spatial scaling filters were used to visualize smaller imaging features, the SD of the pixel distribution histogram and the mean of positive pixels were negatively correlated with nuclear grade on portal venous or nephrographic phase images and were negatively correlated with entropy on unenhanced images. Previous studies of lung cancer showed that uniformity (entropy) was negatively correlated with hypoxia and that the mean of positive pixels was negatively correlated with angiogenesis [24]. It is possible that these features (increased hypoxia, increased angiogenesis, and, in turn, decreased entropy and the mean of positive pixels) may be associated with a higher nuclear grade and may reflect the inverse associations between entropy and the mean of positive pixels seen in the present study. Texture analysis could be complementary to known imaging features (i.e., contrast enhancement characteristics) of RCC on multiphasic CT. However, unlike other imaging biomarkers previously described in relation to RCC, texture analysis can be retrospectively performed on data acquired using standard clinical imaging protocols (including unenhanced CT images). This helps to maximize the information that can be derived from these standard clinical images. Texture analysis is efficient and reproducible. We hope to apply this technique to a larger cohort of patients with RCC, including those with small and medium-sized tumors, to determine whether these associations in particular, associations with clinical outcomes persist. There are some limitations to the present study. The measurements were performed on large renal masses only, because doing so would optimize the chances of obtaining robust texture information. Surgery or full pathologic assessment was eventually performed for all of these large masses. Although texture information can be useful for this cohort of patients with large renal masses (e.g., patients in whom metastatic disease or the presence of sarcomatoid features was noted), further study of small (< 4 cm) and intermediate-sized (4 7 cm) renal masses is warranted, because this type of data would also be very valuable to have for this subgroup of patients. There was some heterogeneity in the CT technique used for the presurgical studies. Therefore, all studies were stratified and analyzed on the basis of the contrast phase, because unenhanced and contrast-enhanced images appear to provide different information. Although most phases were well stratified, data on portal venous and nephrographic phase images were pooled. In addition, noisereducing techniques (e.g., Laplacian of gaussian filtration) are built into the software used in this analysis, thereby limiting the effect of different CT scan parameters. However, a previous study performed using this software has shown that the technology is resistant to differences in acquisition techniques [28]. Use of filtration may reduce sensitivity to enhancement, but there was no built-in correction for that. The present study evaluated pretreatment renal masses only. Some data have suggested that texture analysis could also be used in evaluating response to therapy [26], which was not addressed here but which could be the focus of future research. Finally, only a single slice of the tumor was sampled; volumetric assessment was not performed. There are, however, data suggesting that a single slice may be sufficient for this type of analysis [29, 30]. Not all filter types were available for all masses. The large number of predictors and filters used increases the risk of type I error, so it is possible that some findings of statistical significance may be spurious; however, we corrected for multiple testing by using a Bonferroni correction. In summary, our results reveal that, for patients with large RCCs, texture features (in particular, entropy, the SD of the pixel distribution histogram, and the mean of positive pixels) are associated with histologic type, nuclear grade, and clinical outcomes such as overall survival. In addition, a weak association was also noted between texture measures and the presence of sarcomatoid features and metastatic disease. This type of information could be a useful adjunct to other imaging features and biopsy results, when making management decisions and counseling patients. References 1. Tsao CK, Small AC, Moshier EL, et al. Trends in the use of cytoreductive nephrectomy in the United States. Clin Genitourin Cancer 2012; 10: Minnillo BJ, Zhu H, Maurice MJ, Abouassaly R. Trends in cytoreductive nephrectomy in the eras of immuno and targeted therapy. J Clin Oncol 2014; 32: Abel EJ, Carrasco A, Culp SH, et al. Limitations of preoperative biopsy in patients with metastatic renal cell carcinoma: comparison to surgical pathology in 405 cases. BJU Int 2012; 110: Abel EJ, Culp SH, Matin SF, et al. Percutaneous biopsy of primary tumor in metastatic renal cell carcinoma to predict high risk pathological features: comparison with nephrectomy assessment. J Urol 2010; 184: Heng DY, Wells JC, Rini BI, et al. Cytoreductive nephrectomy in patients with synchronous metastases from renal cell carcinoma: results from the International Metastatic Renal Cell Carcinoma Database Consortium. Eur Urol 2014; 66: Shuch B, Bratslavsky G, Linehan WM, Srinivasan R. Sarcomatoid renal cell carcinoma: a comprehensive review of the biology and current treatment strategies. Oncologist 2012; 17: Vera-Badillo FE, Templeton AJ, Duran I, et al. Systemic therapy for non-clear cell renal cell carcinomas: a systematic review and meta-analysis. Eur Urol 2015; 67: Roubaud G, Gross-Goupil M, Wallerand H, de Clermont H, Dilhuydy MS, Ravaud A. Combination of gemcitabine and doxorubicin in rapidly progressive metastatic renal cell carcinoma and/or sarcomatoid renal cell carcinoma. Oncology 2011; 80: Choi YA, Kim CK, Park SY, Cho SW, Park BK. Subtype differentiation of renal cell carcinoma using diffusion-weighted and blood oxygenation level-dependent MRI. AJR 2014; 203:[web]W78 W Takeuchi M, Kawai T, Suzuki T, et al. MRI for differentiation of renal cell carcinoma with sarcomatoid component from other renal tumor types. Abdom Imaging 2015; 40: Wang W, Ding J, Li Y, et al. Magnetic resonance 104 AJR:207, July 2016

10 CT Texture Features and RCC imaging and computed tomography characteristics of renal cell carcinoma associated with Xp11.2 translocation/tfe3 gene fusion. PLoS One 2014; 9:e Egbert ND, Caoili EM, Cohan RH, et al. Differentiation of papillary renal cell carcinoma subtypes on CT and MRI. AJR 2013; 201: Karlo CA, Di Paolo PL, Chaim J, et al. Radiogenomics of clear cell renal cell carcinoma: associations between CT imaging features and mutations. Radiology 2014; 270: Pierorazio PM, Hyams ES, Tsai S, et al. Multiphasic enhancement patterns of small renal masses ( 4 cm) on preoperative computed tomography: utility for distinguishing subtypes of renal cell carcinoma, angiomyolipoma, and oncocytoma. Urology 2013; 81: Raman SP, Johnson PT, Allaf ME, Netto G, Fishman EK. Chromophobe renal cell carcinoma: multiphase MDCT enhancement patterns and morphologic features. AJR 2013; 201: Young JR, Margolis D, Sauk S, Pantuck AJ, Sayre J, Raman SS. Clear cell renal cell carcinoma: discrimination from other renal cell carcinoma subtypes and oncocytoma at multiphasic multidetector CT. Radiology 2013; 267: Ganeshan B, Panayiotou E, Burnand K, Dizdarevic S, Miles K. Tumour heterogeneity in non-small cell lung carcinoma assessed by CT texture analysis: a potential marker of survival. Eur Radiol 2012; 22: Miles KA, Ganeshan B, Rodriguez-Justo M, et al. Multifunctional imaging signature for V-KI-RAS2 Kirsten rat sarcoma viral oncogene homolog (KRAS) mutations in colorectal cancer. J Nucl Med 2014; 55: Ng F, Ganeshan B, Kozarski R, Miles KA, Goh V. Assessment of primary colorectal cancer heterogeneity by using whole-tumor texture analysis: contrast-enhanced CT texture as a biomarker of 5-year survival. Radiology 2013; 266: Yip C, Landau D, Kozarski R, et al. Primary esophageal cancer: heterogeneity as potential prognostic biomarker in patients treated with definitive chemotherapy and radiation therapy. Radiology 2014; 270: Zhang H, Graham CM, Elci O, et al. Locally advanced squamous cell carcinoma of the head and neck: CT texture and histogram analysis allow independent prediction of overall survival in patients treated with induction chemotherapy. Radiology 2013; 269: Ganeshan B, Miles KA. Quantifying tumour heterogeneity with CT. Cancer Imaging 2013; 13: Ganeshan B, Abaleke S, Young RC, Chatwin CR, Miles KA. Texture analysis of non-small cell lung cancer on unenhanced computed tomography: initial evidence for a relationship with tumour glucose metabolism and stage. Cancer Imaging 2010; 10: Ganeshan B, Goh V, Mandeville HC, Ng QS, Hoskin PJ, Miles KA. Non-small cell lung cancer: histopathologic correlates for texture parameters at CT. Radiology 2013; 266: Miles KA, Ganeshan B, Hayball MP. CT texture analysis using the filtration-histogram method: what do the measurements mean? Cancer Imaging 2013; 13: Goh V, Ganeshan B, Nathan P, Juttla JK, Vinayan A, Miles KA. Assessment of response to tyrosine kinase inhibitors in metastatic renal cell cancer: CT texture as a predictive biomarker. Radiology 2011; 261: Raman SP, Chen Y, Schroeder JL, Huang P, Fishman EK. CT texture analysis of renal masses: pilot study using random forest classification for prediction of pathology. Acad Radiol 2014; 21: Miles KA, Ganeshan B, Griffiths MR, Young RC, Chatwin CR. Colorectal cancer: texture analysis of portal phase hepatic CT images as a potential marker of survival. Radiology 2009; 250: Lubner MG, Stabo N, Lubner SJ, et al. CT textural analysis of hepatic metastatic colorectal cancer: pre-treatment tumor heterogeneity correlates with pathology and clinical outcomes. Abdom Imaging 2015; 40: Ng F, Kozarski R, Ganeshan B, Goh V. Assessment of tumor heterogeneity by CT texture analysis: can the largest cross-sectional area be used as an alternative to whole tumor analysis? Eur J Radiol 2013; 82: Ganeshan B, Miles KA, Young RC, Chatwin CR. Texture analysis in non-contrast enhanced CT: impact of malignancy on texture in apparently disease-free areas of the liver. Eur J Radiol 2009; 70: Gerlinger M, Rowan AJ, Horswell S, et al. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. N Engl J Med 2012; 366: Ba-Ssalamah A, Muin D, Schernthaner R, et al. Texture-based classification of different gastric tumors at contrast-enhanced CT. Eur J Radiol 2013; 82:e537 e Weiss GJ, Ganeshan B, Miles KA, et al. Noninvasive image texture analysis differentiates K-ras mutation from pan-wildtype NSCLC and is prognostic. PLoS One 2014; 9:e Chong Y, Kim JH, Lee HY, et al. Quantitative CT variables enabling response prediction in neoadjuvant therapy with EGFR-TKIs: are they different from those in neoadjuvant concurrent chemoradiotherapy? PLoS One 2014; 9:e Ganeshan B, Skogen K, Pressney I, Coutroubis D, Miles K. Tumour heterogeneity in oesophageal cancer assessed by CT texture analysis: preliminary evidence of an association with tumour metabolism, stage, and survival. Clin Radiol 2012; 67: Mattonen SA, Palma DA, Haasbeek CJ, Senan S, Ward AD. Early prediction of tumor recurrence based on CT texture changes after stereotactic ablative radiotherapy (SABR) for lung cancer. Med Phys 2014; 41: Yip C, Davnall F, Kozarski R, et al. Assessment of changes in tumor heterogeneity following neoadjuvant chemotherapy in primary esophageal cancer. Dis Esophagus 2015; 28: Ganeshan B, Ziauddin Z, Goh V, et al. Quantitative imaging biomarkers from PET-CT as potential correlates for angiogenesis and hypoxia in colorectal cancer. (abstract) Insights Imaging 2012; 3(suppl 1):S Bézy-Wendling J, Kretowski M, Rolland Y, Le Bidon W. Toward a better understanding of texture in vascular CT scan simulated images. IEEE Trans Biomed Eng 2001; 48: Zhang J, Lefkowitz RA, Ishill NM, et al. Solid renal cortical tumors: differentiation with CT. Radiology 2007; 244: Jinzaki M, Tanimoto A, Mukai M, et al. Double-phase helical CT of small renal parenchymal neoplasms: correlation with pathologic findings and tumor angiogenesis. J Comput Assist Tomogr 2000; 24: Hakimi AA, Chen YB, Wren J, et al. Clinical and pathologic impact of select chromatin-modulating tumor suppressors in clear cell renal cell carcinoma. Eur Urol 2013; 63: Kapur P, Pena-Llopis S, Christie A, et al. Effects on survival of BAP1 and PBRM1 mutations in sporadic clear-cell renal-cell carcinoma: a retrospective analysis with independent validation. Lancet Oncol 2013; 14: Shuch B, Bratslavsky G, Shih J, et al. Impact of pathological tumour characteristics in patients with sarcomatoid renal cell carcinoma. BJU Int 2012; 109: AJR:207, July