Synthesizing Multi-session Structural MRI using Age and Patch Information
Poster No:
2148
Submission Type:
Abstract Submission
Authors:
Russell Macleod1, Muriel Bruchhage2, Jonathan O'Muircheartaigh3
Institutions:
1King's College London, London, England, 2Stavanger University, Stavanger , Norway, 3King's College London, London, London
First Author:
Co-Author(s):
Introduction:
Techniques such as normative modelling (5) have become popular to model cross-sectional brain development with respect to a continuous variable, typically age. In neurodevelopmental conditions, quantifying individual differences or deviations from neurotypical brain development has been shown to be powerful in understanding heterogeneous conditions like prematurity, genetic syndromes and psychiatric illness. In early life, normative approaches are ideal as brain development is non-linear and, during short windows, rapid (1). Here we propose an extension of normative modelling to whole brain longitudinal development, modelling brain anatomy at later timepoints, given its appearance at an initial timepoint, and with deviations over time analogous to positive or negative "thrive lines" slopes in paediatric growth charts (2). This approach allows the identification of outliers of typical longitudinal development on a whole brain basis.
Methods:
Images from the Brown University Assessment of Myelination and Behavioral development Across Maturation (BAMBAM) study were utilised for modelling and estimation (3). Images were acquired on a 3T Siemens Trio scanner with repetition/echo times varied according to expected head size described in detail in 3. At 12 months, repetition/echo/inversion times were 16/6.9/950ms, collected at a resolution of 1.4x1.4x1.4mm. From the cohort, we selected 195 subjects with 2 longitudinal scans who had age ranges of 68-4542 (scan 1), 128-5476 (scan 2) and 57-2588 days (scan interval). Image pre-processing included N4 bias correction, brain extraction and registration to template space (FSL+ANTs) as in 3.
Synthetic images were generated using a method based on that seen in (4). Voxel-wise Gaussian Process regression (6) was used to predict intensity values at the 2nd timepoint across the whole brain using two models. The 1st used age at scan and sex (age model) and was trained using the 2nd scans only as output. The 2nd used age-at-scan 1, age-at-scan 2 and sex in combination with voxel intensities from a 5x5x5 patch surrounding each voxel at time 1 (patch model), providing spatial priors for later timepoint prediction. 5-fold cross-validation was used to assess accuracy of the estimated models. Median absolute error (MAE) maps were calculated by averaging the difference between each subject's 2nd scan and equivalent synthetic image. To assess utility of the models, Z-score abnormality maps were created by dividing the difference images by subject voxel-wise GP variance outputs.
Synthetic images were generated using a method based on that seen in (4). Voxel-wise Gaussian Process regression (6) was used to predict intensity values at the 2nd timepoint across the whole brain using two models. The 1st used age at scan and sex (age model) and was trained using the 2nd scans only as output. The 2nd used age-at-scan 1, age-at-scan 2 and sex in combination with voxel intensities from a 5x5x5 patch surrounding each voxel at time 1 (patch model), providing spatial priors for later timepoint prediction. 5-fold cross-validation was used to assess accuracy of the estimated models. Median absolute error (MAE) maps were calculated by averaging the difference between each subject's 2nd scan and equivalent synthetic image. To assess utility of the models, Z-score abnormality maps were created by dividing the difference images by subject voxel-wise GP variance outputs.
Results:
Both models generated synthetic images with intensity values close to those seen in the observed scans. Areas of more consistent structure, such as the basal ganglia, are more accurately modelled while more heterogeneous areas, such as the cortical surface, appear smoothed. However, providing spatial priors from time 1, significantly reduced MAE for time 2. Predicted images with abnormality maps overlays Fig 1 and demonstrate the model's ability to detect a range of structural deformations and lesions. MAE images can be seen in Fig 2 and achieved MAE score (averaging over all voxels in the MAE image) of 0.242 for the age model and 0.222 for the patch model.
Conclusions:
Our models can predict T1w structural images at a 2nd timepoint using either subject demographic information or by combining this with patch information from an initial time point. These images infer voxel intensity values from a typical population but, in the patch model, allowing better incorporation of subject specific tissue boundaries and structural heterogeneity. This can be seen by the lower MAE score for the patch model compared to the age model. The Z-score abnormality maps can detect an enlarged ventricle present only at the 2nd scan, an abnormal cerebellum present at both scans and a lesion in the periventricular tissue that is not present or obscured by motion in the 1st scan. Combined, these demonstrate a potential use of patch-based modelling in voxel-based neuroimaging.
Modeling and Analysis Methods:
Bayesian Modeling 2
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Normal Development 1
Keywords:
Computational Neuroscience
Development
Modeling
MRI
STRUCTURAL MRI
1|2Indicates the priority used for review
Provide references using author date format
Cole, T. J. (2022). 'Growth references and standards.' In Human growth and development (pp. 391-422). Academic Press.
Deoni, S. C. (2012). 'Investigating white matter development in infancy and early childhood using myelin water faction and relaxation time mapping.' Neuroimage, 63(3), 1038-1053.
Macleod, R. (2021). 'Detection of Injury and Automated Triage of Preterm Neonatal MRI Using Patch-Based Gaussian Processes.' In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Perinatal Imaging, Placental and Preterm Image Analysis: 3rd International Workshop, UNSURE 2021, and 6th International Workshop, PIPPI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 3 (pp. 231-241). Springer International Publishing.
Marquand, A. (2016). 'Understanding heterogeneity in clinical cohorts using normative models: beyond case-control studies.' Biological psychiatry, 80(7), 552-561.
Williams, C. K. (2006). 'Gaussian processes for machine learning' (Vol. 2, No. 3, p. 4). Cambridge, MA: MIT press.