Bayesian feature selection for radiomics using reliability metrics

Abstract

Introduction: Imaging of tumors is a standard step in diagnosing cancer and making subsequent treatment decisions. The field of radiomics aims to develop imaging based biomarkers using methods rooted in artificial intelligence applied to medical imaging. However, a challenging aspect of developing predictive models for clinical use is that many quantitative features derived from image data exhibit instability or lack of reproducibility across different imaging systems or image-processing pipelines. Methods: To address this challenge, we propose a Bayesian sparse modeling approach for image classification based on radiomic features, where the inclusion of more reliable features is favored via a probit prior formulation. Results: We verify through simulation studies that this approach can improve feature selection and prediction given correct prior information. Finally, we illustrate the method with an application to the classification of head and neck cancer patients by human papillomavirus status, using as our prior information a reliability metric quantifying feature stability across different imaging systems.

Keywords: Bayesian modeling; classification; probit prior; quantitative imaging; radiomics; variable selection.

FIGURE 1

Overview of radiomics. Here, we use images from Aerts et al. (2014) to illustrate the initial step of tumor imaging and segmentation. Quantitative features can then be extracted that summarize the tumor boundary (shape features), the distribution of the pixel intensity values within the tumor (histogram features), and aspects of the spatial relations among pixels with differing intensities (texture features). More advanced features such as model- or tree-based summaries may also be computed. These features can then be used as input to approaches aimed at prediction of survival outcomes or classification methods such as mixture models, as illustrated generically in Step 3.

FIGURE 2

Schematic illustration of sparse Bayesian classification model. Feature-specific prior information is incorporated through the probit prior.

FIGURE 3

The posterior probability of inclusion for all variables from the proposed method, for one simulated data set. The first four variables are the discriminatory variables, and the remaining 100 are the noise variables.

FIGURE 4

Classification accuracy: ROC curves for the RVS model, the Bayesian model with a neutral prior, lasso logistic regression, and SVM on an example simulated test set, as described in Section 4.1.2.

FIGURE 5

Classification accuracy: ROC curves for the RVS model, the Bayesian model with a neutral prior, lasso logistic regression, and SVM on a test set with unequal group sizes, as described in Section 4.1.3.