Presentation Authors: Felix Yap*, Aliasger Shakir, Bino Varghese, Steven Cen, Darryl Hwang, Christopher Lau, Lindsay Yang, Derek Liu, Marielena Rivas, Passant Mohamed, Xiaomeng Lei, Manju Aron, Mihir Desai, Inderbir Gill, Vinay Duddalwar, Los Angeles, CA
Introduction: Differentiating benign from malignant masses using enhancement alone can be difficult. Additional imaging metrics (e.g. tumor shape and texture) have been shown to improve discrimination. Using a computer algorithm to quantitatively analyze shape and texture features of renal tumors in three dimensions on routine imaging, we created an objective non-invasive method to evaluate tumor behavior. We tested this radiomics framework in objectively distinguishing between benign and malignant renal masses on imaging. We also determined whether shape shows independency from texture in discriminating benign from malignant masses.
Methods: Multiphase computed tomography (CT) images of 135 patients with 94 malignant and 41 benign masses diagnosed between 2011 and 2014 were segmented. Point coordinates of tumor contours in all axial slices were input into a MATLAB (MathWorks) algorithm. 33 shape metrics and 760 texture metrics were calculated per tumor. We used least absolute shrinkage and selection operator (LASSO) to nominate important shape and texture metrics respectively in discriminating benign and malignant masses, and we used lambda value to constrain the model size under 4 predictors. Area under the curve (AUC) was used to assess discrimination power. We conducted Z test between shape and texture metrics to investigate the independency of shape and texture from each other. SAS 9.4 was used for all data analysis.
Results: We identified sagittal elliptic compactness and coronal Feret ratio as the best shape metrics, and minimal signal and entropy in corticomedullary and excretory phases respectively as the best texture metrics. AUC using both shape and texture (0.88) was significantly greater than the AUCs using only shape (0.64) and texture (0.82) alone (p = 0.05). Moreover, there was statistically significant difference between AUCs for shape metrics and texture metrics (p = 0.01).
Conclusions: Shape is an independent predictor in distinguishing benign from malignant tumors. In order to distinguish benign from malignant renal masses, combining both shape and texture metrics in a single radiomics panel outperforms shape and texture metrics in isolation from each other. Future studies will determine further clinical applications of our evolving panel and its potential for better understanding of natural history, behavior, and growth of renal tumors.
Source of Funding: Whittier Foundation