A Fast and Effective Strategy for Artifact Identification and Signal Restoring with HARDI data
Elisa Scaccianoce1,2, Francesca Baglio2, Giuseppe Baselli1, and Flavio Dell'Acqua3

1Department of Electrinics, Informations and Bioengineering, Politecnico di Milano, Milano, Italy, 2RM Lab, Don Carlo Gnocchi Foundation ONLUS, IRCCS S. Maria Nascente, Milano, Milano, Italy, 3NATBRAINLAB, Department of Neuroimaging, Institute of Psychiatry, Psychology and Neuroscience, King’s College, London, United Kingdom, London, United Kingdom

Synopsis

HARDI datasets are often prone to different type of artifacts, difficult to detect even by expert users. In this work we propose a fast and effective pipeline for outlier identification and correction of HARDI datasets. Here corrupted data is first identified as outlier and then regenerated using a framework based on signal decomposition using spherical harmonics. This approach was tested on healthy controls and validated with simulated dataset. Our study confirms the efficacy of using SH for artifacts identification and correction.

Background and purpose

High Angular Resolution Diffusion Imaging (HARDI) is currently considered one of the most suitable approach for study white matter connectivity and is routinely adopted in several clinical research studies. However, with HARDI, a big amount of raw data is also collected, making difficult to manually inspect all of Diffusion Weighted (DW) volumes for artefacts that may negatively impact further analyses. These artifacts, due to head movements, physiological noise and scanner-related issue, are often also subtle and arduous to detect visually, even by expert users. Tools for artifacts identification, based on the evaluation of outliers when computing the residual of the signal fitted by different models[1,2], have been already developed. However, these methods detect outliers at the individual voxel level while most, if not all, artifacts affecting HARDI data are usually impacting more voxels within the same DW volume and brain slice. Moreover, most of this approaches simply reject outliers and fit the specific diffusion model with the remaining data points making potentially this estimation more computational demanding (i.e. applying a customized b-matrix or fiber response model per each voxel). In this work we adopted a fast slice-wise approach to detect corrupted slices within different DW volumes and regenerate the missing data using spherical harmonics (SH) decomposition of the HARDI signal. SH decomposition has also per se an intrinsic beneficial effect of denoising and removal of any not antipodal symmetric features of the HARDI signal. The final output is a fully regenerated dataset that can be then processed with any existing diffusion pipeline.

Methods

HARDI data from 15 healthy subjects, mean age 32 ± 5 years, were acquired using a 3T GE HDx system (General Electric, Milwaukee, WI, USA) with the following parameters: voxel size 2.4x2.4x2.4 mm, slices 60, b-value 3000 s/mm2, 60 diffusion-weighted directions and 7 non-diffusion weighted volumes. After correcting for motion and eddy current distortions using FSL (fsl.fmrib.ox.ac.uk/fsl), for each subject the following steps were performed using a custom written Matlab code (R2013a, www.mathworks.com): 1) HARDI data was fitted using a SH decomposition. 2) Residuals values were computed as the difference between the actual HARDI signal and the SH modelled signal. For each brain slice the mean value of this residual was calculated along each DW direction. 3) A binary outlier mask was created by identifying as outliers all DW directions with a mean residual value above an automatic threshold obtained for each slice across all DW directions and defined as: $$$threshold=Q3+IQR\times3$$$, where Q3 is the 75th percentile and IQR is the interquartile range computed on the mean residuals[3]. 4) Corrupted signals were regenerated using new SH coefficients obtained by SH decomposition performed this time without outlier directions. SH at order 6 was used when regenerating slices of corrupted directions while SH at order 8 was applied to the rest of the data.

Results

To validate our method, a simulated artifact was created in a subject who did not present any artifact. In figure 1, panel a) shows the corrupted dataset in the sagittal view while panel b) displays the actual (left) and the regenerated (right) data. On real data, visual inspection of the diffusion dataset identified 2 subject with obvious and significant artefacts of along different DW directions. All artifacts were detected and corrected by the proposed approach. In figures 2 and 3, two examples of artifact correction are shown. Figure 2a illustrates the outlier mask in which one of the outlier value is identified and its location displayed in sagittal and axial views before (Fig. 2b) and after (Fig. 2c) correction. Similar results are show in figure 3. Few seconds were required to correct each subject on normal laptop.

Discussion and Conclusion

This work presents an effective and fast tool for the automatic outlier detection and correction. Our study confirms the efficacy of using SH for artifacts identification and proposes a completely model independent approach and simple solution for their correction. This approach can be useful in large studies when visual quality check may be not practical or when data artefacts are not obvious at visual inspection. To conclude, the present work introduces a fully automatic method for both artifact identification and correction of HARDI data which can potentially increase the precision in DW-derived measures.

Acknowledgements

No acknowledgement found.

References

[1] Pannek, K., Raffelt, D., Bell, C., Mathias, J. L., & Rose, S. E. (2012). HOMOR: Higher order model outlier rejection for high b-value MR diffusion data. Neuroimage, 63(2), 835-842.

[2] Chang, L., Walker, L., & Pierpaoli, C. (2012). Informed RESTORE: A method for robust estimation of diffusion tensor from low redundancy datasets in the presence of physiological noise artifacts. Magnetic Resonance in Medicine, 68(5), 1654-1663.

[3] Leemans A., Evans C.J., Jones D.K. (2008). Quality assessment through analysis of residuals of diffusion image fitting. Proceedings of the 16th ISMRM, Toronto.

Figures

Figure1. Panel a) shows the corrupted dataset in the sagittal and axial view while panel b) displays the actual (left) and the regenerated (right) data.

Figure2. Panel a illustrates the outlier mask in which one of the outlier value is identified and its location displayed in sagittal and axial views before (panel b) and after (panel c) correction.

Figure3. Panel a illustrates the outlier mask in which one of the outlier value is identified and its location displayed in sagittal and axial views before (panel b) and after (panel c) correction.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
2081