Purpose: Blind source separation algorithms such as independent component analysis (ICA) and non-negative matrix factorization (NMF) can identify functional modes from fMRI scans. Results from NMF are easier to interpret than ICA, but NMF assumes that input is non-negative whereas fMRI pre-processing typically produces values with zero-mean and unit-variance to identify functional signal change. Different studies have employed various strategies for dealing with these negative values. The goal of this work was to evaluate the impact of four strategies in a test-retest fMRI dataset.
Methods: A public dataset was downloaded from openneuro.org/datasets/ds000114. It consisted of 10 subjects scanned twice, spaced 2-3 days apart, while performing five behavioral tasks. All images were spatially aligned and then temporally concatenated for each session and task, resulting in 10 pre-processed series (2 sessions x 5 tasks). The fMRI data was standardized to zero-mean and unit-variance for ICA while four strategies were applied to remove negativity for NMF by: (1) setting negatives to zero, (2) taking absolute value, (3) using spectra magnitude after Fourier Transform, or (4) using signal magnitude without any standardization. Sixteen spatial components were computed for each task using ICA and NMF and the agreement between the two scanning sessions was evaluated using the dice coefficient.
Results: Among the 80 computed components (16 components x 5 tasks), the number of NMF components with a dice coefficient above 0.5 was 39 if the data was zeroed, 27 if absolute values were used, 11 if the spectra magnitude were used, and 29 if the signal magnitude was used. For ICA, 28 components had a dice score above 0.5.
Conclusion: The strategy for removing negativity has substantial impact on NMF performance on fMRI detection. This study showed that the best approach was to set negative values to zero after data standardization.