Biologically Constrained Augmentation for Optimal Segmentation of MRI with Ventricular Abnormality
Poster No:
2002
Submission Type:
Abstract Submission
Authors:
Mohamad Zeina1, Linda D'Antona2, Robert Gray1, Mikael Brudfors3, Guilherme Pombo1, Amy Nelson1, Suraj Sennik2, Sophie Mullins2, Laura Santos2, Manjit Matharu1, Laurence Watkins2, Ahmed Toma2, Parashkev Nachev1
Institutions:
1UCL Queen Square Institute of Neurology, London, United Kingdom, 2The National Hospital for Neurology and Neurosurgery, London, United Kingdom, 3NVIDIA Corporation, London, United Kingdom
First Author:
Co-Author(s):
Introduction:
Tissue segmentation of brains with marked ventricular abnormalities is complicated by the heterogeneity of the associated patterns of anatomical distortion, obscuring underlying structure-function relationships. Conventional segmentation models employing standard augmentation techniques exhibit brittleness to ventricular morphology, and this can only be definitively corrected by representations of the space of possible ventricular appearances. A principled approach to augmenting standard segmentation models is to simulate, in a biologically plausible manner, the effects of ventricular pathology across its full range.
Methods:
In this study, we develop a novel approach to biologically plausible augmentation of MR brain images based on topologically constrained deformation of the ventricular compartment. We utilised two distinct head MRI datasets: the publicly accessible OASIS-3 (Marcus et al., 2007), without examples of cerebrospinal fluid (CSF) flow disorders and that could be reliably segmented by traditional techniques, and a local dataset from UCL Queen Square Institute of Neurology, encompassing a variety of CSF flow disorders and that could not be reliably segmented using traditional techniques.
A deep learning model was trained to produce augmented brains. The input to the model is the ventricle class from OASIS-3 brains segmented using Geodesic Information Flows (Cardoso et al., 2015), and the output is a deformation field that is applied to the input to produce an augmented ventricle segmentation. The model is trained by optimising a three-part objective: to maximise, within constraints, the relative growth in ventricle size, to minimise overall deformation magnitude, and to maintain spatial smoothness across the deformation field.
This model was then used to produce a set of augmented scans (Fig 1) with corresponding precise tissue class segmentation maps. A 3D U-Net model (Ronneberger et al., 2015) was trained on a) only the unaltered scans and b) a combination of unaltered and altered scans. Some local scans were used for training both models, in cases where traditional segmentation approaches were successful. Trained models were evaluated against OASIS-3 scans, with and without augmentation, that were held out during training.
A deep learning model was trained to produce augmented brains. The input to the model is the ventricle class from OASIS-3 brains segmented using Geodesic Information Flows (Cardoso et al., 2015), and the output is a deformation field that is applied to the input to produce an augmented ventricle segmentation. The model is trained by optimising a three-part objective: to maximise, within constraints, the relative growth in ventricle size, to minimise overall deformation magnitude, and to maintain spatial smoothness across the deformation field.
This model was then used to produce a set of augmented scans (Fig 1) with corresponding precise tissue class segmentation maps. A 3D U-Net model (Ronneberger et al., 2015) was trained on a) only the unaltered scans and b) a combination of unaltered and altered scans. Some local scans were used for training both models, in cases where traditional segmentation approaches were successful. Trained models were evaluated against OASIS-3 scans, with and without augmentation, that were held out during training.
Results:
The deep learning model trained on this augmented dataset exhibited superior segmentation accuracy when evaluated on the most challenging unseen brains, reflected in higher median dice scores for all 3 tissue classes (Grey Matter: 0.651, White Matter: 0.801, CSF: 0.693), compared to a model trained exclusively on unaltered scans (Grey Matter: 0.577, White Matter: 0.684, CSF: 0.666), shown in Fig 2. The improvement persisted across brains at every level of ventricle enlargement (Mean dice 0.700 for all classes when trained with augmentation and 0.6740 without). This enhancement underscores the potential of our deep learning-based augmentation pipeline in significantly improving performance for segmenting anatomically atypical brains. Moreover, a subset of the augmented brains exhibited a biologically plausible appearance, similar to that in CSF flow disorders, shown in Fig 1. This further validates the effectiveness of our approach in generating synthetic, realistic representations of rare anatomical variations.
Conclusions:
The enhanced accuracy of tissue class segmentation for brains with CSF flow disorders, when using our deep learning augmentation pipeline, exemplifies the crucial role that deep learning has in benefiting patients with a broad range of anatomical abnormalities that are typically underrepresented in conventional datasets. Our approach effectively bridges the gap between the scarcity of representative data and the need for precise segmentation. This success suggests that more comprehensive generative models, capable of simulating a broader range of rare imaging variations, hold great promise for future advancements in neuroimaging and personalised medicine generally.
Modeling and Analysis Methods:
Image Registration and Computational Anatomy
Methods Development
Segmentation and Parcellation 1
Other Methods
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Neuroanatomy Other 2
Keywords:
Cerebro Spinal Fluid (CSF)
Computational Neuroscience
Computing
Data analysis
Machine Learning
Segmentation
Spatial Warping
STRUCTURAL MRI
Structures
Workflows
1|2Indicates the priority used for review
Provide references using author date format
Marcus, D. S. et al. (2007), 'Open Access Series of Imaging Studies (OASIS): Cross-sectional MRI Data in Young, Middle Aged, Nondemented, and Demented Older Adults', Journal of Cognitive Neuroscience, 19(9), 1498–1507. https://doi.org/10.1162/jocn.2007.19.9.1498
Ronneberger, et al. (2015), 'U-Net: Convolutional Networks for Biomedical Image Segmentation', In N. Navab, J. Hornegger, W. M. Wells, & A. F. Frangi (Eds.), Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 (pp. 234–241). Springer International Publishing. https://doi.org/10.1007/978-3-319-24574-4_28