An Improved Optimization Method for the Relevance Voxel Machine

摘要

In this paper, we will re-visit the Relevance Voxel Machine (RVoxM), a recently developed sparse Bayesian framework used for predicting biological markers, e.g., presence of disease, from high-dimensional image data, e.g., brain MRI volumes. The proposed improvement, called IRVoxM, mitigates the shortcomings of the greedy optimization scheme of the original RVoxM algorithm by exploiting the form of the marginal likelihood function. In addition, it allows voxels to be added and deleted from the model during the optimization. In our experiments we show that IRVoxM outperforms RVoxM on synthetic data, achieving a better training cost and test root mean square error while yielding sparser models. We further evaluated IRVoxM's performance on real brain MRI scans from the OASIS data set, and observed the same behavior - IRVoxM retains good prediction performance while yielding much sparser models than RVoxM.