A Theoretical Prediction of CT-based Radiomics Feature Variance to Improve the Performance of Classification Models.
Authors
Affiliations (4)
Affiliations (4)
- Department of Bioengineering at the University of Pennsylvania.
- Biomedical Engineering at Johns Hopkins University.
- Department of Radiology at the University of California Los Angeles.
- Department of Radiology at the University of Pennsylvania.
Abstract
<i>Radiomics</i> involves the high-throughput extraction of quantitative image biomarkers for analysis in machine learning models. While this field holds immense potential for personalized medicine, the clinical translation of radiomic models is severely hampered by their intrinsic dependency on image acquisition properties, degrading downstream predictive performance. In this work, we propose a statistical framework to explicitly account for stochastic-based feature uncertainty. We derive an analytic variance estimator through a novel use of noise power spectrum (NPS) information to linearly propagate image noise to radiomic features. The accuracy of this estimator was validated against Monte Carlo simulations. The predicted and measured feature variance were compared through z-score normalized percentage error with the predicted variance achieving a prediction error of <20% across the majority of histogram and gray level co-occurrence features. We subsequently integrated these variance estimates to an uncertainty aware support vector machine (TSVC [1]) and compare the results to the standard support vector machine model. We simulated a simple classification problem with known per-sample variance estimates and show that TSVC has the capacity to either perform just as well, or better than the standard SVM. These results demonstrate that 1) variance estimates of features can be reliably obtained and 2) illustrate the potential utility of incorporating variance information to enhance model performance.