Optimizing WMH Segmentation for diverse clinical datasets with SynthSegCSVD
Poster No:
2005
Submission Type:
Abstract Submission
Authors:
erin gibson1, Joel Ramirez1, Lauren Woods1, Stephanie Berberian1, Julie Ottoy2, Christopher Scott1, Fuqiang Gao1, Roberto Coello3, maria Hernandez3, Anthony Lang4, Carmela Tartaglia5, Malcolm Binns6, Robert Bartha7, sean symons5, Richard Swartz1, Mario Masellis1, Alan Moody1, Bradley MacIntosh1, Joanna Wardlaw3, Sandra Black1, Andrew Lim1, Maged Goubran5
Institutions:
1Sunnybrook Health Sciences Centre, Toronto, Canada, 2Sunnybrook Health Sciences Centre, University of Toronto, Toronto, Canada, 3University of Edinburgh, Edinburgh, United Kingdom, 4Toronto Western Hospital, Toronto, Canada, 5University of Toronto, Toronto, Canada, 6Baycrest Hospital, Toronto, Canada, 7University of Western Hospital, London, Canada
First Author:
Co-Author(s):
Introduction:
White matter hyperintensities (WMH) are key imaging biomarkers of cerebral small vessel disease (CSVD) on FLAIR MRI scans and are associated with a range of worse clinical outcomes, including increased risk for stroke and dementia. Automated, reliable WMH segmentation is crucial but challenging due to data heterogeneity across imaging protocols and scanner hardware. The segmentation task is further complicated by the diverse characteristics of WMH. Existing tools often fail to generalize across varied imaging datasets. This study presents SynthSegCSVD, an advanced CNN-based tool with a UNet architecture designed for improved WMH segmentation in heterogenous clinical datasets with varying degrees of CSVD burden.
Methods:
SynthSegCSVD was developed using a large dataset consisting of over 1000 scans sourced from seven multi-site studies, encompassing a range of clinical populations, WMH burdens, and imaging protocols. A novel two-stage segmentation framework was developed that first leverages FreeSurfer's SynthSeg (Billot, 2023) to generate a targeted regional mask containing two key neuroanatomical structures, and subsequently combines this mask with the FLAIR image for improved WMH segmentation. Advanced machine learning strategies, including the ensembling of three models with distinct precision-recall weightings and test-time augmentation, were utilized to ensure robust segmentation performance. The efficacy of SynthSegCSVD was evaluated by benchmarking its performance against two state-of-the-art segmentation tools, HyperMapper (Forooshani, 2022) and SAMSEG (Cerri, 2023), using several diverse test datasets.
Results:
SynthSegCSVD exhibited superior segmentation performance across all test datasets, surpassing the benchmark tools in both accuracy and reliability (Fig. 1). Its superior performance was most evident in datasets that employed isotropic FLAIR acquisition protocols, where a significant reduction in WMH contrast was also observed (permuted p<0.001). In this more challenging segmentation scenario, SynthSegCSVD demonstrated a significant increase in the mean Dice score compared to HyperMapper and SAMSEG, with improvements of 0.19 and 0.34, respectively (permuted p<0.001). Furthermore, SynthSegCSVD exhibited robustness to variations in image orientation and header inaccuracies, maintaining remarkable stability across a range of imaging conditions and patient populations.
Conclusions:
SynthSegCSVD represents a significant advancement in automated WMH segmentation, effectively addressing the challenges of data heterogeneity. Its robust performance across a broad spectrum of imaging conditions and patient characteristics makes it a promising tool for large-scale studies and clinical applications, particularly in populations with varying degrees of CSVD.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 2
Modeling and Analysis Methods:
Segmentation and Parcellation 1
Keywords:
Cerebrovascular Disease
MRI
Open-Source Software
Segmentation
White Matter
1|2Indicates the priority used for review
Provide references using author date format
Cerri S. (2023). 'An open-source tool for longitudinal whole-brain and white matter lesion segmentation'. Neuroimage: Clinical, vol. 38, pp. 1-17.
Forooshani P. (2022). 'Deep Bayesian networks for uncertainty estimation and adversarial resistance of white matter hyperintensity segmentation'. Human Brain Mapping, vol. 43(7), pp. 2089-2108.