Vinay Anant Duddalwar MD, FRCR, Abstract Co-Author: Research Grant, General Electric Company

Bhushan Desai MBBS, MS, Presenter: Nothing to Disclose

Darryl Hwang PhD, Abstract Co-Author: Nothing to Disclose

Steven Cen PhD, Abstract Co-Author: Nothing to Disclose

Frank K. Chen MD, Abstract Co-Author: Nothing to Disclose

Hannu Tapio Huhdanpaa MD, Abstract Co-Author: Nothing to Disclose

Phillip Ming-Da Cheng MD, MS, Abstract Co-Author: Nothing to Disclose

Inderbir Gill MD, Abstract Co-Author: Nothing to Disclose

To build a prediction model using quantitative imaging metrics (QIM) derived from contrast enhanced computed tomography (CECT) to distinguish clear cell renal cell carcinoma (ccRCC) from papillary RCC (pRCC).

We retrospectively queried the surgical database and found 72 post nephrectomy patients who had pathology proven ccRCC (53) or pRCC (19) and preoperative multiphase CECT of the abdomen. Voxel-based contrast enhancement values were collected from the lesion segmentation and displayed as a histogram. Mean and median enhancement and histogram distribution parameters skewness, kurtosis, standard deviation (SD), and interquartile range (IQR) were calculated for each lesion on corticomedullary phase.  Independent t-test was used for normally distributed parameters while Wilcoxon rank sum test was used for not normally distributed parameters. Supervised machine learning (Classification and Regression Tree 7.0-CART®) was used to develop the prediction model.

ccRCC had significantly higher mean and median whole lesion enhancement, IQR and SD (p < 0.01), and significantly lower skewness and kurtosis (p < 0.01) compared to pRCC. Arterial mean and venous IQR were selected as the final predictors. ROC curve showed by using these two factors the model can reach the accuracy of AUC=0.89 (95% CI: 0.81, 0.96). The cut points selected by CART are: if arterial mean > 75 Hounsfield Units (HU) or arterial mean ≤ 75HU and venous IQR ≤ 301HU then the lesion will be classified as cRCC. Otherwise, if arterial mean ≤ 75HU and venous IQR > 301HU then the lesion is pRCC. From the learning sample only, this prediction rule reached 88.7% sensitivity and 94.7% specificity. When we applied a 10-fold cross validation, the estimated generalizable sensitivity and specificity are 77.4% and 73.7% respectively.

A prediction model encompassing QIM seems promising and can be used as a quantitative tool to differentiate ccRCC from pRCC. Further refinements with possible inclusion of additional QIM (spherocity, lobularity of lesion) and validation on an independent dataset are currently underway.

The successful integration and validation of novel imaging-based biomarker methodologies (such as QIM) will improve our ability to stratify patients at risk, increase diagnostic accuracy, help establish guidelines for active surveillance in the management of RCC and optimize criteria used for clinical decision making.

Duddalwar, V, Desai, B, Hwang, D, Cen, S, Chen, F, Huhdanpaa, H, Cheng, P, Gill, I, Development of a Clinical Prediction Model Based on Quantitative Imaging Metrics Derived from CT Imaging for Differentiating Clear Cell from Papillary Renal Cell Carcinoma.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14016677.html