Decoding the encoding of functional brain networks: An fMRI classification comparison of non-negative matrix factorization (NMF), independent component analysis (ICA), and sparse coding algorithms

- We compared ICA, K-SVD, NMF, and L1-Regularized Learning for encoding brain components within an fMRI scan.
- The temporal weights of each encoding were used to predict activity using machine learning classifiers.
- NMF, which eliminates negative BOLD signal, performed poorly compared to ICA and sparse coding algorithms (K-SVD, L1 Regularized Learning).
- L1 Regularized Learning and K-SVD frequently outperformed four variations of ICA to predict fMRI task activity.
- Spatial sparsity of encoding maps were associated with increased classification accuracy, holding constant effects of algorithms.

BackgroundBrain networks in fMRI are typically identified using spatial independent component analysis (ICA), yet other mathematical constraints provide alternate biologically-plausible frameworks for generating brain networks. Non-negative matrix factorization (NMF) would suppress negative BOLD signal by enforcing positivity. Spatial sparse coding algorithms (L1 Regularized Learning and K-SVD) would impose local specialization and a discouragement of multitasking, where the total observed activity in a single voxel originates from a restricted number of possible brain networks.New methodThe assumptions of independence, positivity, and sparsity to encode task-related brain networks are compared; the resulting brain networks within scan for different constraints are used as basis functions to encode observed functional activity. These encodings are then decoded using machine learning, by using the time series weights to predict within scan whether a subject is viewing a video, listening to an audio cue, or at rest, in 304 fMRI scans from 51 subjects.Results and comparison with existing methodThe sparse coding algorithm of L1 Regularized Learning outperformed 4 variations of ICA (p < 0.001) for predicting the task being performed within each scan using artifact-cleaned components. The NMF algorithms, which suppressed negative BOLD signal, had the poorest accuracy compared to the ICA and sparse coding algorithms. Holding constant the effect of the extraction algorithm, encodings using sparser spatial networks (containing more zero-valued voxels) had higher classification accuracy (p < 0.001). Lower classification accuracy occurred when the extracted spatial maps contained more CSF regions (p < 0.001).ConclusionThe success of sparse coding algorithms suggests that algorithms which enforce sparsity, discourage multitasking, and promote local specialization may capture better the underlying source processes than those which allow inexhaustible local processes such as ICA. Negative BOLD signal may capture task-related activations.

99Visual network manually identified across each algorithm, within a single scan. Sparsifying algorithms (K-SVD and LASSO/L1-Regularization) outperformed ICA and NMF algorithms for predicting whether a subject was viewing a video, listening to an audio stimulus, or resting, during an fMRI scan. Maps were rescaled to be on common scale for illustration purposes.