Abstract
290
Objectives: Promising results for metabolic connectivity as obtained by sparse inverse covariance estimation (SICE) of 18F-FDG PET data have been reported. However, this statistical method lacks robustness in the connectivity estimation. Metabolic connectivity matrices were observed to present similar patterns as structural connectivity matrices obtained from diffusion MRI. Given this similarity, we propose to regularize the sparse estimation of metabolic connectivity using structural connectivity.
Methods: In contrast to a single regularization parameter in the conventional SICE [1], the structural weighting introduces an additional matrix regularization parameter that is based on structural connectivity from DTI tractography. Thus, it applies an individualized penalty to any potential metabolic connection. Five different ways of optimizing the regularization matrix were examined. The method was first validated on simulated FDG PET data with predefined metabolic connectivity. Afterwards, the method was applied to a clinical data set of patients with Alzheimer’s disease (AD, n=41), frontotemporal lobar degeneration (FTLD, n=30) and healthy subjects (n=26), who were imaged on a hybrid PET/MR system. The classification was based on the probability estimated from group-wise multivariate Gaussian distributions.
Results: In the simulation study (ROC curves in the figure), the structure-weighted SICE outperformed the conventional SICE in recovering the predefined metabolic connections for almost all the explored structural weighting schemes. The structural weighting is insensitive to the penalty weight selection and is thus more stable than conventional SICE. The possible false negative estimation of the structural weighting only happens at very weak metabolic connections. Given a very high baseline performance of the conventional SICE (accuracy of 93-96%), the proposed method provided only a minor improvement in the differentiation between patients and healthy subjects (1-2 % higher accuracy). However, it increased more than 5% accuracy of the differentiation between AD and FTLD, which is a major clinical interest.
Conclusion: According to the results of both the simulation and clinical data, the structural weighting improves stability in the estimation of metabolic connectivity. This study further underlines an added value of multi-modal imaging as provided by hybrid PET/MR systems. Research Support: No