Statistical Clustering of Parametric Maps from Quantitative Dynamic Contrast Enhanced MRI and an Associated Decision Tree Model for Non-Invasive Tumor Grading of Solid Clear Cell Renal Cell Carcinoma
Yin Xi1, Qing Yuan1, Yue Zhang1, Ananth Madhuranthakam1,2, Jeffrey Cadeddu3, Vitaly Margulis3, James Brugarolas4,5, Payal Kapur3,6, and Ivan Pedrosa1,2

1Radiology, UTSouthwestern Medical Center, Dallas, TX, United States, 2Advanced Imaging Research Center, UTSouthwestern Medical Center, Dallas, TX, United States, 3Urology, UTSouthwestern Medical Center, Dallas, TX, United States, 4Internal Medicine, UTSouthwestern Medical Center, Dallas, TX, United States, 5Developmental Biology, UTSouthwestern Medical Center, Dallas, TX, United States, 6Pathology, UTSouthwestern Medical Center, Dallas, TX, United States


We propose a method that provides a simplified visual representation of tumor vascular heterogeneity in clear cell renal cell carcinoma (ccRCC) based on the combination of multiple parametric maps from quantitative dynamic contrast-enhanced (DCE) MRI analysis. Using this approach we observed an association between the tumor grade and vascular heterogeneity, especially in medium size tumors. A decision tree model was developed to predict high grade and low grade histology in solid ccRCCs, which may help in management decisions by providing additional information about the tumor biology beyond tumor size.


Grading of renal cell carcinoma (RCC) can help in the decision between surgery and active surveillance, and is an important determinant of disease prognosis.[1] Current methods for pre-surgical tumor grading rely on percutaneous biopsies, which are invasive and prone to sampling errors in heterogeneous tumors. A magnetic resonance imaging (MRI)-based non-invasive method for grading would be attractive because it explores tumor heterogeneity in the whole tumor. Clear cell RCC (ccRCC), the most common renal malignancy, is well suited to approaches that assess tumor vascularity due to its known dependency on angiogenesis. A comprehensive quantitative analysis of tumor vascularity can be achieved using dynamic contrast enhanced (DCE) MRI with high temporal resolution and subsequent analysis of these imaging datasets using pharmacokinetic models such as the one proposed by Tofts et al. (1). While some studies have analyzed the ability of DCE-derived parametric maps for grading RCC[2], to our knowledge the value of combining the information provided by these maps into a single, composite representation of tumor vascular heterogeneity has not been reported. This study has two purposes. First, we employ a statistical clustering method to partition each tumor into different areas associated with their level of enhancement on DCE MRI. Second, we use a decision tree model to predict low and high grade histology.


Patients and MRI Protocol: This was a prospective, IRB-approved, HIPAA-compliant study. After signing an informed consent, 42 patients scheduled for surgical resection of a known solid renal mass later proven to represent a ccRCC underwent 3T MRI with a phased array Torso coil (Achieva or Ingenia, Philips Medical System, Cleveland, OH). DCE MRI was performed using a coronal 3D spoiled gradient echo acquisition with a temporal resolution of 5 sec. To minimize respiratory motion, three consecutive dynamic phases were obtained within each 15-sec breath-hold. A 15-sec period of free-breathing was allowed between consecutive acquisition periods. Three baseline dynamic phases were acquired, followed by a bolus of 0.1 mmol/kg of gadobutrol (Gadavist; Bayer Healthcare Pharmaceuticals, Wayne, NJ) using a power injector at a rate of 2 cc/sec followed by a 20 cc saline flush at 2 cc/sec. Image Analysis: DCE images were analyzed using commercial software VersaVue Enterprise (iCAD Inc., Nashua, NH). Pixel-by-pixel fitting with Tofts model was performed after motion correction to generate quantitative maps of Ktrans and Kep. The initial area under the concentration curve (iAUC) was also calculated. All DCE-derived maps were then analyzed with a DICOM viewer (OsiriX). Regions of interest (ROI) of entire tumor were drawn on the DCE-derived maps Ktrans, Kep, and iAUC. Histopathology: The histopathological analysis of the tumor after surgical resection served as the gold-standard. All tumors were classified as: (a) low-grade (LG) ccRCC for Fuhrman grades 1 and 2; (b) high-grade (HG) ccRCC for Fuhrman grades 3 and 4. Statistics: First, pixel maps for all tumors were obtained. The length of the tumor (cm) was calculated based on the tumor area assuming a spherical shape. Based on the DCE values, fuzzy c-means (FCM) clustering method was used to cluster all pixels into three categories, namely low-active area (LAA), median-active area (MAA) and high-active area (HAA). The percentages of the areas within the tumor were also computed. Next, these percentages together with the length of the tumor were considered as predictors for a decision tree model. In order to reduce the impact of overfitting, leave-one-out cross validation was implemented while calculating accuracy, sensitivity, specificity, positive predictive value and negative predictive value.


Seventeen high grade and 25 low grade ccRCCs were included in this study. Mean tumor size was 20.4 cm2 ± 17.9 cm2 which corresponds to a mean largest diameter of 3.3 cm ± 1.4 cm. DCE maps from all patients were successfully obtained. FCM method was able to distinguish different tumor areas according to their DCE enhancement. High grade histology was associated with larger %HAA in medium size tumors (n=21, p = 0.023, Figure 1). Images from two representatives tumors are shown in Figure 2. An optimal decision tree model was constructed and is plotted in Figure 3. After adjusting by cross validation, the optimal tree reached an accuracy of 74%, with 88% sensitivity, 64% specificity, 63% PPV, and 89% NPV for the diagnosis of high grade ccRCC.


The FCM offers an opportunity to summarize multiple DCE-derived pharmacokinetic maps and identify unique tumor regions with different enhancement characteristics. Using this approach we were able to construct a decision tree model using criteria beyond size to predict tumor grade, especially in medium length tumors.


Supported by Grant NIH/NCI 1R01CA154475


1. P. S. Tofts, et al., JMRI 10-3: 223–232, 1999.

2. Sun, et al. Radiology 250, 3, 2009.


Scatter-plot of percentage of high-active area (%HAA) against tumor length (cm). While small tumors are predominantly low grade and larger tumors are high grade, prediction of grade in medium length tumors (shaded area) is challenging. Among them, high grade tumors have higher %HAA than low grade ones (p = 0.023).

MRI in 2 ccRCCs. From left: T2-W; DCE maps (iAUC, Kep, Ktrans); and FCM. Tumor 2 is smaller than Tumor 1 (2.9 cm vs. 3.6 cm), but has higher %HAA (55%, red on FCM) suggestive of HG, which was confirmed at histopathology. Tumor 1 (%HAA = 10%) was LG ccRCC.

Flow chart for classifying high grade (HG) and low grade (LG) solid ccRCC tumors based on optimal decision tree model.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)