Transformer-based unified tissue and white matter hyperintensity segmentation
Poster No:
154
Submission Type:
Abstract Submission
Authors:
James Ruffle1, Guilherme Pombo1, Chris Foulon2, Robert Gray1, Samia Mohinta1, Holger Engleitner1, Harpreet Hyare1, Geraint Rees1, Parashkev Nachev1
Institutions:
1UCL Queen Square Institute of Neurology, London, London, 2UCL Brain Sciences, London, London
First Author:
Co-Author(s):
Introduction:
The causes and consequences of white matter hyperintensities (WMH), amongst the commonest findings in neuroradiology, remain a subject of intense study. UK Biobank provides the largest research cohort of MRI studies with accompanying WMH labels, with great potential to cast light on this important condition. Unfortunately, many of the lesion labels are spurious, misidentifying healthy choroid plexus, normal ependymal lining, or imaging artefact, leading to reduced accuracy and potential confounding of downstream analytic tasks. Here we derive a substantially improved set of WMH labels through the combination of classical and vision-transformer based segmentation methods.
Methods:
In a cohort of 33241 UK Biobank participants, we develop a pipeline employing classical multi-channel volumetric (T1 and FLAIR) tissue segmentation, brain extraction, and WMH segmentation based on a vision transformer model with a custom tissue-leveraging loss. We quantify segmentation fidelity out-of-sample with conventional metrics, intra-subject stability over time, and a downstream predictive task.
Results:
We obtained out-of-sample Dice scores for brain extraction (>0.999), gray matter (0.98), white matter (0.99), cerebrospinal fluid (0.96), and WMH (0.87). At 8 seconds per participant, processing time was significantly faster than conventional tissue segmentation with SPM (218 seconds, p<0.0001). No differences in model performance were observed between young or old, and male or female (p>0.05), indicating good demographic calibration. WMH segmentation showed significantly greater temporal stability compared with current UK Biobank WMH labels (p<0.0001) for a subset of the cohort imaged twice. Age regression based on WMH count was superior compared with the original labels (R2=0.18 vs 0.14) (Figure 1).
Conclusions:
We provide a tissue and WMH segmentation model optimized for UK Biobank data with evidence of fidelity superior to existing labels. We make our model and segmentations available through UK Biobank to assist in the study of this important disorder.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Lifespan Development:
Aging
Modeling and Analysis Methods:
Segmentation and Parcellation 2
Neuroinformatics and Data Sharing:
Databasing and Data Sharing
Keywords:
Aging
Cerebrovascular Disease
Degenerative Disease
Machine Learning
Modeling
Segmentation
STRUCTURAL MRI
White Matter
1|2Indicates the priority used for review
·Figure 1
Provide references using author date format
PyTorch - https://pytorch.org