Combining lesion burden with cortical malformation morphology strongly predicts motor outcomes in children with cerebral palsy
Alex Pagnozzi1, Nicholas Dowson1, James Doecke1, Simona Fiori2, Andrea Guzzetta3, Roslyn N Boyd4, and Stephen Rose1

1The Australian e-Health Research Centre, CSIRO Health & Biosecurity, Brisbane, Australia, 2Stella Maris Institute, Pisa, Italy, 3Stella Maris institute, Pisa, Italy, 4The University of Queensland, Queensland Cerebral Palsy and Rehabilitation Research Centre, Brisbane, Australia

Synopsis

Magnetic Resonance Imaging (MRI) is the clinical standard for assessing developmental brain injury in children with Cerebral Palsy (CP). We propose an automated process that segments the spectrum of white and grey matter injury, including tissue lesions and malformations of the cortex, and correlates biomarkers of injury with the Assisting Hand Assessment (AHA), a clinical score quantifying hand function. The proposed method is shown to perform accurate tissue and injury segmentation using T1 and T2 MRI compared to the manual classification of injury, and was significantly correlated with AHA (p<0.001).

Purpose

To identify key imaging biomarkers related to tissue lesions and cortical malformations from the MRIs of children with CP, and to develop data-driven structure-function relationships that can help predict motor impairment for children with CP.

Methods

Our hypothesis is we can identify biomarkers of injury from structural MRI which are strongly predictive of motor outcomes. Therefore a robust brain tissue segmentation method is presented, tailored to severely injured cases observed in children with CP. This segmentation approach is based on the Expectation Maximisation (EM)/Markov Random Field (MRF) approach (Van Leemput et al., 1999; Zhang et al., 2001), with a modification to weight clique potentials within the MRF based on image intensities and initialised with a peak-finding algorithm on the image histogram, in order to minimise reliance on atlas priors, which introduces significant errors in patients with severe injury. These modifications were implemented in the MATLAB R2015b programming environment (Mathworks, Natick, MA), and applied to a cohort of 211 children aged 5-17, including 167 children with unilateral CP and 44 children with healthy development, who were all scanned using high-resolution structural MPRAGE at 3T. On a subset of 23 cases, 17 with severe injury, the proposed segmentation method demonstrated improved Dice Similarity Coefficients (DSCs) on both healthy data (0.83 vs 0.81) and cases with severe cortical injury (0.78 vs 0.76) compared to state of the art software including FreeSurfer and ANT’s Atropos, as shown for in a specific example in Fig 1. Measures of cortical thickness and the depth of sulci from the inner surface of the skull were computed from the cortical segmentation using the Laplacian method, while cortical curvature was computed using the Visualisation Toolkit (VTK). Cortical regions were segmented using the Colin 27 Automated Anatomically Labelling (AAL) atlas using level sets to propagate the atlas labels. Cortical shape measures were converted to a z-score compared to the healthy mean and variance of these measures in the respective cortical regions, and used as biomarkers of cortical malformations. The z-score of cortical thickness in each cortical region, averaged across all children with CP is illustrated in Fig 2. Tissue lesions were identified using lesion belief maps constructed from the healthy white and grey matter, and cerebrospinal fluid distributions elucidated using the EM-weighted MRF segmentation, registered tissue probability maps, and the aligned T2-weighted MRIs. Refinement of the lesion segmentation was then performed using the EM algorithm. The regional involvement of the grey and white matter regions were computed using the AAL atlas and the ICBM DTI-81 Atlas (International Consortium for Brain Mapping, CA) respectively, and used as biomarkers for focal lesions. The frequency of observed lesion involvement between grey and white matter regions in this cohort of children with CP is presented in Fig 3.

Results and Discussion

Multivariable regression models were constructed from a 50% training set to identify associations between cortical shape alone, lesion involvement alone, and cortical shape and lesion involvement combined, with the AHA score. Data-driven variable selection was performed using the ‘stepAIC’ package in R Statistical Software to retain only the most predictive cortical and lesion involvement biomarkers for each regression model. The cortical shape only model had an R-squared of 0.88 on the training set, retaining features on the lingual gyrus, gyrus rectus and middle temporal gyrus (p<0.001), and performed well in the 50% validation test set (Pearson’s r = 0.62, p<0.001). The lesion involvement only model had an R-squared of 0.42 on the training set, retaining solely the lenticular nucleus as a significant predictor (p=0.001), and similarly performed well in the test set (Pearson’s r = 0.53, p<0.001). The combined model achieved an R-squared of 0.94 on the training data, retaining thickness of the primary motor cortex, supplementary motor area, primary somatosensory cortex, as well the lenticular nucleus and cerebral peduncle as significant predictors of AHA (all p<0.05). All of these significant predictors have a known role in motor function, suggesting that true structure-function relationships of the brain have been elucidated. This model obtained further improvements to performance on the independent test set (Pearson’s r = 0.66, p<0.001), as shown in Fig 4, compared to cortical shape or lesion involvement alone, highlighting that the two biomarkers of injury capture separate portions of variance in motor outcome.

Conclusion

The improved model performance when combining measures of cortical shape and lesion involvement highlights the importance of characterising both types of injury, while the generalisability of the trained models to unseen data demonstrates the ability to predict motor outcomes for children with CP.

Acknowledgements

Alex M. Pagnozzi is supported by the Australian Postgraduate Award (APA) from The University of Queensland, and the Commonwealth Scientific Industrial and Research Organisation (CSIRO). Roslyn N. Boyd is supported by a Foundation for Children Grant, NHMRC Career Development Fellowship (1037220) and a NHMRC Project Grant COMBIT (1003887). Roslyn N. Boyd and Stephen Rose are supported by the by the Smart Futures Co-Investment Program Grant. The funding bodies have not contributed to the study design, the collection, management, analysis and interpretation of data, the writing of final reports or the decision to submit findings for publication. No other authors have potential conflicts of interest to declare.

References

Van Leemput, K., Maes, F., Vandermeulen, D., Suetens, P., 1999. Automated model-based tissue classification of MR images of the brain. IEEE Trans. Med. Imaging 18, 897–908. doi:10.1109/42.811270 Zhang, Y., Brady, M., Smith, S., 2001. Segmentation of brain MR images through a hidden Markov random field model and the expectation-maximization algorithm. IEEE Trans. Med. Imaging 20, 45–57. doi:10.1109/42.906424

Figures

Fig 1 Cortical grey matter segmentations from a manual expert (cyan), the modified segmentation algorithm (yellow), FreeSurfer (red) and ANT’s Atropos (green).

Fig 2 The average z-score of cortical thickness for children with CP, compared to the healthy population, for each cortical region.

Fig 3 Frequency of lesion involvement in grey and white matter regions.

Fig 4 Performance of the trained combined regression model on the independent test set.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
0733