A Hybrid LDA+gCCA Model for fMRI Data Classification and Visualisation.

IEEE Trans Med Imaging. 2014 Nov 26;

Authors: Afshin-Pour B, Shams SM, Strother S

Abstract
Linear predictive models are applied to functional MRI (fMRI) data to estimate linear boundaries that can discriminate scans acquired during different experimental task states. These boundaries, (i.e. visualised as discriminant statistical parametric maps (SPMs)) provide significantly non-random predictive power, and range from low to high spatial reproducibility across subjects (e.g., [1, 2]). Such inter-subject pattern reproducibility is an essential characteristic of interpretable SPMs that generalise across subjects. It also reduces the likelihood of missing nodes in spatial networks that present temporally more variable blood oxygenation level dependant (BOLD) signal, and weak association with the experimental task [3]. Therefore, we introduce a flexible model that optimises spatial pattern reliability by simultaneously enhancing the prediction power and reproducibility of linear predictive models. Maximizing the prediction enhances detection of the task-positive brain network, and maximizing spatial reproducibility facilitates detection of more temporally variable, spatially reproducible networks. This hybrid model is formed by a weighted summation of the optimization functions of a linear discriminate analysis (LDA) model and a generalized canonical correlation (gCCA) model [4]. The linear discriminate term preserves the model's ability to discriminate the fMRI scans of multiple brain states while gCCA finds a linear combination for each subject's scans such that the estimated boundary map is reproducible across subjects. The hybrid model is implemented in the NPAIRS (i.e. Nonparametric, Prediction, Activation, Influence, Reproducibility, re-Sampling) split-half resampling framework and compared with a Fisher's discriminant analysis [5]. In the NPAIRS framework, the available data are randomly split into two sets, and on each set a single model is fitted and an SPM is estimated. The NPAIRS split-half resampling framework provides reproducibility (r) and prediction (p) metrics for that model. The LDA, the proposed hybrid, and Gaussian Naive Bayes (GNB) techniques were applied to simulated fMRI data, and the results show that the hybrid model outperforms the other two techniques in terms of receiver operating characteristic (ROC) curves, particularly for detecting less predictable but spatially reproducible networks and/or nodes. These techniques were applied to real fMRI data to estimate the maps for two task contrasts. For both task contrasts our results indicate that compared to LDA, and GNB the hybrid model can provide maps with large increases in reproducibility for small reductions in prediction, which are jointly closer to the ideal performance point of (p=1, r=1). The results from simulated data and real fMRI data confirm that the hybrid model provides a better reflection of the interaction of the default mode network, and task positive network in a single activation map.

PMID: 25438304 [PubMed - as supplied by publisher]