Program Information
Predicting Gene Mutations in Renal Cell Carcinoma Using Machine Learning
D Staub*, R Hannan , K Thomas , S Jiang , I Pedrosa , P Kapur , J Brugarolas , J Wang , UT Southwestern Medical Center, Dallas, TX
Presentations
TU-AB-BRA-2 (Tuesday, July 14, 2015) 7:30 AM - 9:30 AM Room: Ballroom A
Purpose: The goal of this project is to investigate the use of software extracted features from contrast-enhanced CT to predict the genetic mutations commonly present in renal tumors using machine learning.
Methods: Our initial IRB approved patient database consisted of 33 patients with renal cell carcinoma (RCC). For each patient, the database contained a contrast-enhanced CT image and physician delineated contour outlining the RCC tumor boundary in the CT image. In addition, tumor tissue from each patient was biopsied to determine the presence of gene mutations in BAP1, VHL, and PBRM1. Features based on tumor geometry, intensity, and texture were extracted from the contrast-enhanced CT image of each patient. Features were used to train a support vector machine (SVM) classifier to predict expression of each gene separately. Hyperparameter grid search and feature selection meta-algorithms coupled with cross-validation were employed to protect against overfitting of the SVM model.
Results: The average cross-validation accuracy was used to evaluate the predictive model. Average accuracy was 0.87, 0.91, and 0.9 for VHL, BAP1, and PBRM1 respectively. Texture features were the most prominent feature type in the models for all three genes.
Conclusion: Using our models we observed predictive accuracy >87% for all three gene mutations evaluated. Accurate predictive models could allow medical images to serve as convenient surrogates for expensive and time consuming gene assay procedures.
Contact Email: