Payel Ghosh, Presenter: Nothing to Disclose

Arvind Rao, Abstract Co-Author: Nothing to Disclose

Pheroze Tamboli MD, Abstract Co-Author: Nothing to Disclose

Raghunandan Vikram MBBS, FRCR, Abstract Co-Author: Nothing to Disclose

To identify image-derived features on contrast-enhanced CT scans in patients with Clear Cell Renal Carcinoma (CRCC) that are predictive of BAP1 deletion.

We performed a feasibility study on contrast enhanced CT scans from a set of 39 patients (11 with BAP1 deletion) from The Cancer Genome Atlas (TCGA) KIRC database. We created tumor segmentation masks using the open-source software, Medical Image Interaction Tool Kit (mitk.org). Three-dimensional textural features (such as Laws’ textural features, Wavelet transforms, and Haralick texture measures), volumetric features, and ratios of features at different image resolutions were computed using MATLAB. A total of 73684 different imaging features were extracted for each phase of the CT (non-contrast (nc), cortico-medullary (cm), nephrographic(neph) and excretory(ex)), and correlated to the BAP1 loss (derived from the TCGA database). Feature association with mutation status was tested using the Wilcoxon rank-sum test. Multiple testing correction for p-values was done using Benjamini-Hochberg FDR correction. A statistical random forest classifier was trained on features significant at an unadjusted p-value of 0.05 to predict mutation status. Receiver Operating Characteristics (ROC curves) were obtained using a five-fold cross-validation procedure.

The number of features that were significantly different in tumors with BAP1 loss were: 3994 in the non-contrast phase; 475 in the cortico-medulary phase; 1844 in the nephrographic phase; and 3907 in the excretory phase. Out of these features, nephrographic and excretory phases had 182 and 273 features that were significantly different after FDR correction (q-value < 0.05). Area Under the Curve (AUC) values from the classifier based on the features (from unadjusted p-value < 0.05) identified nephrographic and non-contrast phases as reliable predictors of BAP1 loss (AUC:1, and AUC:0.8 respectively).

Our initial feasibility study suggests that it is possible to predict BAP1 status from analysis of imaging features derived from contrast enhanced CT. Nephrographic and non-contrast phases seem to be most predictive in this initial analysis.

BAP I deletion is associated with high grade CRCC and poor prognosis. We demonstrate a non-invasive way of predicting this on contrast enhanced CT scans.

Ghosh, P, Rao, A, Tamboli, P, Vikram, R, Image Features Derived from Contrast Enhanced CT Can Predict BAP1 Loss in Patients with Clear Cell Renal Carcinoma: Results of a Feasibility Study.  Radiological Society of North America 2013 Scientific Assembly and Annual Meeting, December 1 - December 6, 2013 ,Chicago IL. http://archive.rsna.org/2013/13044402.html