hide
Free keywords:
-
Abstract:
Introduction
Resting-state fMRI signal has been coupled to brain-wide neuronal signal fluctuations, presenting large-scale integration1,2,3. Spontaneous signals display <0.1 Hz oscillatory patterns at varied brain states4. Recently, a single-vessel fMRI method has been developed to characterize the rs-fMRI fluctuation of individual vessels5,6. Here, we hypothesize that the spectral feature of the slow oscillatory pattern can be learnt by an artificial neural network. We used Echo State Networks7 (ESN) to encode single-vessel BOLD fluctuations and predict the <0.1 Hz slow temporal dynamics 10 seconds ahead.
Methods
Data from 6 rats were acquired with a 14.1T magnet using the bSSFP8 method. 6 adult subjects were scanned using an EPI sequence in a 3T scanner. In both cases one slice TR was 1 s and the duration of each trial was 15 minutes.
Using the A-V map5 and ICA9 single-vessel time courses were extracted only from veins exhibiting a strong slow fluctuation. After normalizing the data, the signals have been bandpass filtered in either 0.01-0.05 Hz (rat) or 0.01-0.1 Hz (human) frequency ranges to extract the slowly changing feature.
ESN, an artificial neural network belonging to the class of reservoir computing methods, has been used to encode the slow oscillations using supervised learning (Fig 1A).
Surrogate data10 served as controls verifying the degree of encoding performed by the chosen ESNs.
Results/Discussion
To encode the slow spontaneous dynamics, ESNs have been trained to predict the extracted features shifted by 10 seconds with regard to the normalized raw data. The networks were trained separately for human and rat data. ESN's hyperparameters have been optimized using random search11 and their performances have been evaluated by computing Pearson correlation coefficients between network-predictive outputs and input fresh data (Fig.1A,2B). In both human and rat cases the predictions of real data obtained significantly higher scores than those of controls (Fig.1CG,2B). The optimized ESN reservoir trained on 4 trials of one rat was used to predict slow fMRI signal fluctuations of other rats (Fig.1FH). Also, an ESN trained on human vessels has been employed to forecast time courses extracted from V1 ROIs of 250 subjects obtained from the Human Connectome Project12 data. The results demonstrate that it is possible to distinguish brain states based on the obtained prediction scores (Fig.2FG)
Conclusions
Using high-resolution imaging methods allowed us to extract data from individual venules and target a specific biological mechanism. We have shown that the vascular spontaneous slow oscillations are in many cases predictable and that vessels across subjects or specimen share common oscillatory features. By predicting V1 fluctuations of HCP subjects we demonstrated the connection between vascular and whole-brain dynamics. This paves the way for encoding intrinsic activity of the entire brain.
The low computational demands of the method make it a good fit for real-time neurofeedback applications.
