5466

Predicting Individual Task Performance From Resting State fMRI: Effects of Training Task Data Quality

Alexander D. Cohen¹, Elizabeth Zakszewski¹, and Yang Wang¹

¹Radiology, Medical College of Wisconsin, Milwaukee, WI, United States

Synopsis

Resting state functional MRI (rs-fMRI) has been used to predict individual task activation by training a model to map rs-fMRI networks to task performance. This study used a multiband, multi-echo acquisition to collect motor task fMRI as training data. The effects of echo combination and denoising of the training-task data on rs-fMRI predictions were examined. Multi-echo task data resulted in increased predictive accuracy of the model. These results suggest the quality of the training-task data affects the accuracy of the prediction model.

Introduction

Resting state functional MRI (rs-fMRI) has been used to predict task activation on an individual basis by training a linear-regression based machine learning model^1,2. These studies have used a wide variety of tasks with varying results. Language tasks, for example, tend to perform better than motor tasks in terms of the predictive accuracy of the resting state scans². In this study, a motor task was performed during a multiband (MB), multi-echo (ME) acquisition³ to examine the effect of the task acquisition on the prediction accuracy of the model. Specifically, the impact of echo combination and denoising using multi-echo independent component analysis (MEICA)⁴ was examined.

Methods

Fourteen subjects were imaged on a 3T system. Eight subjects returned within two weeks for repeat scans. High-resolution anatomical images (MPRAGE) were collected for co-registration with the functional images.

Eyes-closed MB rs-fMRI data was collected with parameters as follows: TR/TE=802/33.5ms, FOV=208mm, 2mm isotropic resolution, MB acceleration=8, 72 slices, FA=50°, blipped-CAIPI FOV-shift=3⁵, 498 repetitions. Following the rs-fMRI acquisition, a one-minute scan was collected with reversed phase encoding direction. Each subject also performed a task consisting of four alternating 40s periods of rest and bilateral finger tapping. Scans were acquirewasd using a MB, ME simultaneous ASL/BOLD scan3 with: TR=4.0s, TE=9.1,25,39.6,54.3ms, FOV=240mm, 3mm isotropic resolution, MB acceleration=4, in-plane R=2, 36 slices, FA=90°, blipped-CAIPI FOV shift=3, pCASL labeling.

Preprocessing for rs-fMRI and scans included image distortion correction using topup^6,7, volume registration, denoising with FIX^8,9, coregistration to MNI space, and spatial smoothing with a 4.5mm FWHM Gaussian kernel. Task-fMRI processing focused on the BOLD component of the signal. The processing scheme is shown in Figure 1. Two datasets were analyzed. The second echo (E2) was processed to mimic typical single-echo fMRI acquisitions. In addition, data was processed following echo combination and denoising using the multi-echo independent component analysis (MEICA) technique (MECDN), which removes non-BOLD signal from the data^4,10,11. A GLM was used to extract task activation and ASL oscillations were regressed form the data.

Subjects’ predicted motor task maps were derived based on Parker-Jones et al¹. First, the preprocessed rs-fMRI timeseries and task activation maps were converted to cifti format. Features were then extracted from the rs-fMRI data using a dual-regression analysis on a set of 33 group features derived from the ICA of 100 Human Connectome Project (HCP)^12,13 subjects, resulting in individual feature maps. A linear regression model was trained to map from individual feature maps to task activation. The resulting beta coefficients were averaged with a “leave-one-out” analysis for each subject to create a predicted task activation map. Pearson correlation (PC) and dice coefficients (DC) were calculated between each subject’s predicted and all other subjects’ task maps for the E2 and MECDN datasets. Matrices were row- and column-normalized and evaluated for diagonality. The repeatability of the predicted activation was calculated for subjects with multiple scans as the DC between predicted activation at TP1 and TP2 and compared between E2 and MECDN scans.

Results

Overall, the MECDN data more accurately predicted activation on an individual basis compared to the E2 data. This is evidenced by the stronger diagonal component on the MECDN correlation matrices (Figure 2), which indicates each subject’s predicted activation map correlates more with their own actual activation mask vs. the other subjects’. A significant increase in the mean of the diagonal elements was observed between the MECDN and E2 datasets for PC (2.13+/-0.14 vs. 1.73+/-0.17, P=0.003), and DC (1.83+/-0.20 vs. 1.06+/-0.22, P=0.01) matrices. The mean of the PC and DC diagonal elements for both datasets were significantly higher than zero (P<0.0001). Example overlap maps between predicted and actual activation are shown in Figure 3. Overall, repeatability was high (DC>0.75). No differences in repeatability were seen between the datasets.

Discussion and Conclusions

This study found that predicted task activation from rs-fMRI benefited from a ME acquisition and subsequent MEICA denoising. A motor task was chosen because these tasks have shown less accuracy in predictions in previous studies². Using a ME acquisition resulted in more accurate and repeatable task predictions. This likely stems from the increased SNR from the inherent averaging in the echo combination and the denoising procedure. Thus, while ME acquisitions lead to an increased TR, the reduction in time points is outweighed by the SNR increases. In addition, only 14 subjects were imaged in this study. More subjects should increase the robustness of the predictions. Future studies could examine the results with a more traditional MBME fMRI approach. Overall, these results indicate the task fMRI data used to train a predictive model using rs-fMRI affects the predictive accuracy of the model.

Acknowledgements

This work was supported by a Daniel M. Soref Charitable Trust Grant (to Y.W.).

References

1. Parker Jones O, Voets NL, Adcock JE, Stacey R, Jbabdi S. Resting connectivity predicts task activation in pre-surgical populations. NeuroImage Clinical 2017;13:378-385.

2. Tavor I, Parker Jones O, Mars RB, Smith SM, Behrens TE, Jbabdi S. Task-free MRI predicts individual differences in brain activity during task performance. Science 2016;352(6282):216-220.

3. Cohen AD, Nencka AS, Wang Y. A Novel Multiband Multi-Echo Simultaneous ASL/BOLD Sequence for Task-Based fMRI. 2017; Honolulu, HI. p 3832.

4. Kundu P, Inati SJ, Evans JW, Luh WM, Bandettini PA. Differentiating BOLD and non-BOLD signals in fMRI time series using multi-echo EPI. Neuroimage 2012;60(3):1759-1770.

5. Setsompop K, Gagoski BA, Polimeni JR, Witzel T, Wedeen VJ, Wald LL. Blipped-controlled aliasing in parallel imaging for simultaneous multislice echo planar imaging with reduced g-factor penalty. Magn Reson Med 2012;67(5):1210-1224.

6. Smith SM, Jenkinson M, Woolrich MW, Beckmann CF, Behrens TE, Johansen-Berg H, Bannister PR, De Luca M, Drobnjak I, Flitney DE, Niazy RK, Saunders J, Vickers J, Zhang Y, De Stefano N, Brady JM, Matthews PM. Advances in functional and structural MR image analysis and implementation as FSL. Neuroimage 2004;23 Suppl 1:S208-219.

7. Andersson JL, Skare S, Ashburner J. How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging. Neuroimage 2003;20(2):870-888.

8. Griffanti L, Salimi-Khorshidi G, Beckmann CF, Auerbach EJ, Douaud G, Sexton CE, Zsoldos E, Ebmeier KP, Filippini N, Mackay CE, Moeller S, Xu J, Yacoub E, Baselli G, Ugurbil K, Miller KL, Smith SM. ICA-based artefact removal and accelerated fMRI acquisition for improved resting state network imaging. Neuroimage 2014;95:232-247.

9. Salimi-Khorshidi G, Douaud G, Beckmann CF, Glasser MF, Griffanti L, Smith SM. Automatic denoising of functional MRI data: combining independent component analysis and hierarchical fusion of classifiers. Neuroimage 2014;90:449-468.

10. Kundu P, Brenowitz ND, Voon V, Worbe Y, Vertes PE, Inati SJ, Saad ZS, Bandettini PA, Bullmore ET. Integrated strategy for improving functional connectivity mapping using multiecho fMRI. Proc Natl Acad Sci U S A 2013;110(40):16187-16192.

11. Evans JW, Kundu P, Horovitz SG, Bandettini PA. Separating slow BOLD from non-BOLD baseline drifts using multi-echo fMRI. NeuroImage 2015;105:189-197.

12. Glasser MF, Sotiropoulos SN, Wilson JA, Coalson TS, Fischl B, Andersson JL, Xu J, Jbabdi S, Webster M, Polimeni JR, Van Essen DC, Jenkinson M. The minimal preprocessing pipelines for the Human Connectome Project. NeuroImage 2013;80(0):105-124.

13. Barch DM, Burgess GC, Harms MP, Petersen SE, Schlaggar BL, Corbetta M, Glasser MF, Curtiss S, Dixit S, Feldt C, Nolan D, Bryant E, Hartley T, Footer O, Bjork JM, Poldrack R, Smith S, Johansen-Berg H, Snyder AZ, Van Essen DC, Consortium WU-MH. Function in the human connectome: task-fMRI and individual differences in behavior. NeuroImage 2013;80:169-189.

Figures

Schematic showing the processing steps for the E2 and MECDN datasets. For the MECDN data, each echo was despiked. Then, the first echo was volume registered, registered to the anatomical T1 image, and transformed to MNI space. All other echoes were registered using the transformation matrices from the first echo. Echoes were combined using a T2*-weighted technique, and then non-BOLD components were regressed from the data using MEICA. For the MECDN data, the GLM did not include detrending or regression of motion and tag/control oscillations, which were removed by MEICA.

The correlation coefficient (top) and dice coefficient (bottom) between the actual and predicted task activation for E2 (left) and MECDN (right) task fMRI scans. All matrices are row- and column-normalized. All matrices show the characteristic diagonal trend, however, the trend appears stronger for the MECDN data. Each row/column represents a separate scanning session. Some subjects were scanned twice. These repeat scans were placed in consecutive rows/columns in the matrices. These scans show high correlation with each other as identified by "boxes" of high correlation. For example, row/column 3 and 4 are the same subject scanned twice.

Example activation overlap maps for three representative subjects showing the actual activation (black), predicted activation (red) and the overlap of the two (yellow). In general, more overall activation and activation overlap was observed when the MECDN task data was used to train the model (right) compared to when the E2 data was used to train the model (left).

Proc. Intl. Soc. Mag. Reson. Med. 26 (2018)

5466