Toward Precise Localization of Abnormal Brain Activity: 1D CNN on Single Voxel fMRI Time-Series

Abstract

Functional magnetic resonance imaging (fMRI) is one of the best techniques for precise localization of abnormal brain activity non-invasively. Machine-learning approaches have been widely used in neuroimaging studies; however, few studies have investigated the single-voxel modeling of fMRI data under cognitive tasks. We proposed a hybrid one-dimensional (1D) convolutional neural network (1D-CNN) based on the temporal dynamics of single-voxel fMRI time-series and successfully differentiated two continuous task states, namely, self-initiated (SI) and visually guided (VG) motor tasks. First, 25 activation peaks were identified from the contrast maps of SI and VG tasks in a blocked design. Then, the fMRI time-series of each peak voxel was transformed into a temporal-frequency domain by using continuous wavelet transform across a broader frequency range (0.003-0.313 Hz, with a step of 0.01 Hz). The transformed time-series was inputted into a 1D-CNN model for the binary classification of SI and VG continuous tasks. Compared with the univariate analysis, e.g., amplitude of low-frequency fluctuation (ALFF) at each frequency band, including, wavelet-ALFF, the 1D-CNN model highly outperformed wavelet-ALFF, with more efficient decoding models [46% of 800 models showing area under the curve (AUC) > 0.61] and higher decoding accuracies (94% of the efficient models), especially on the high-frequency bands (>0.1 Hz). Moreover, our results also demonstrated the advantages of wavelet decompositions over the original fMRI series by showing higher decoding performance on all peak voxels. Overall, this study suggests a great potential of single-voxel analysis using 1D-CNN and wavelet transformation of fMRI series with continuous, naturalistic, steady-state task design or resting-state design. It opens new avenues to precise localization of abnormal brain activity and fMRI-guided precision brain stimulation therapy.

Keywords: 1D-CNN; continuous task states; fMRI; single-voxel analysis; wavelet transformation.

Figure 1

Task functional magnetic resonance imaging (fMRI) designs. (A) fMRI-blocked design for finger movement task. Each functional run includes three visual-guided (VG) task blocks, and three self-initiated (SI) task blocks, plus six rest blocks. In VG block, a visual cue of “finger” image was shown on the screen for every 2 s, indicating the time for the participant to press the button. In SI block, a clock sign was shown on the screen during the entire block. (B) fMRI continuous-state design. Participants performed each of the two tasks (i.e., SI and VG) for the entire functional run, each lasting for 8 min. The order of tasks is balanced between participants.

Figure 2

Pipeline of hybrid 1D convolutional neural network (1D-CNN) prediction model. We first extracted the blood-oxygen-level-dependent (BOLD) signal from a specified target voxel (e.g., Occiptial_Mid_L: left middle occipital gyrus) for each subject under each condition (i.e., SI or VG). We then applied continuous wavelet transforms (CWTs) to the time-series to generate time-frequency decomposition across continuous scales. The transformed time-series of each frequency band were then imported to the 1D-CNN model to predict the continuous state.

Figure 3

Linear correlation of between the area under the curve (AUC) of wavelet amplitude of low frequency fluctuation (wavelet-ALFF) and the |T| value of the paired t-test between two continuous task states (i.e., SI vs. VG).

Figure 4

Differences in brain activations between SI and VG tasks in the blocked design [Gaussian random field (GRF) corrected with voxel p < 0.001 and cluster p < 0.05]. Warm colors (red) indicate higher brain activation or activity in the SI condition than VG; cool colors (blue) indicate higher brain activation or activity in the VG condition than SI. The Z-coordinates were from −30 to +65 with a step of 5 mm. Left in the figure is the left in the brain.

Figure 5

The AUC of 25 peak voxels. The names of each peak voxel were listed in Table 1. (A) Decoding performance of 1D-CNN models on the transformed time-series across 32 frequency bands as well as on the original functional magnetic resonance imaging (fMRI) time-series. (B) Decoding performance of wavelet-ALFF across 32 frequency bands. The read dashed lines indicate the threshold of AUC > 0.61 (corresponding to paired |T| > 4.08 and p < 0.0001). The red crosses in A indicate the AUCs of original fMRI time-series, with a mean rank of 20.4 across the 25 voxels, i.e., below the average performance of 1D-CNN. Both the 1D-CNN and wavelet-ALFF models showed the highest decoding performance in the 13th peak (Occipital_Mid_L in Table 1) and the second best performance in the 1st peak (Cuneus_L in Table 1). (C) Receiver operating characteristic (ROC) of 1D-CNN on two exemplar peak voxels, i.e., the 13th peak (Occipital_Mid_L, in the left panel) and the 1st peak (Cuneus_L, in the right panel). Red lines in the plots indicate the ROC of 1D-CNN on wavelet-transformed time-series at a specific frequency. Note that different frequencies were chosen for the two peak voxels according to their best performance, namely, 0.233 and 0.083 Hz, respectively. Blue lines and orange lines indicate the ROC of wavelet-ALFF and original fMRI time-series, respectively.

Figure 6

The relationship of decoding performance in the 1D-CNN and wavelet-ALFF models. We used area under the curve (AUC) to evaluate the decoding of continuous task states for both methods and only the models with AUC > 0.61, corresponding to |T| > 4.08 (p < 0.0001) in the paired t-test on wavelet-ALFF decoding models. There were 800 pairs of decoding models in total across 32 frequency bands and 25 peak voxels. 1D-CNN models showed more efficient predictions of task states than wavelet-ALFF (369 vs. 99 models, respectively) as well as higher decoding accuracies (94% of efficient models).

Figure 7

Distribution of decoding performance across 32 frequency bands for both 1D-CNN (A) and wavelet-ALFF (B). Each dot in the figure represents the AUC results of 1D-CNN (A) or wavelet-ALFF (B) for each peak voxel at each frequency band, in total of 25 (peak voxels) × 32 (frequency bands) = 800 dots. The read dashed lines indicate the efficient predictions with a threshold of AUC > 0.61 (corresponding to |T| > 4.08 and p < 0.0001). The 1D-CNN showed better predictions on high frequency bands (>0.1 Hz), while wavelet-ALFF showed more uniform distribution with a preference on low frequency bands (<0.1 Hz).