Atrophy Assessment via Freesurfer SynthSR: A Clinical Validation
Poster No:
320
Submission Type:
Abstract Submission
Authors:
Severin Schramm1, Melissa Thalhammer2, Benita Schmitz-Koep1, Kirsten Jung1, Dennis Hedderich1
Institutions:
1Department of Diagnostic and Interventional Neuroradiology, School of Medicine, Munich, Bavaria, 2TUM Neuroimaging Center, Munich, Bavaria
First Author:
Severin Schramm, Dr. med. sci
Department of Diagnostic and Interventional Neuroradiology, School of Medicine
Munich, Bavaria
Department of Diagnostic and Interventional Neuroradiology, School of Medicine
Munich, Bavaria
Co-Author(s):
Benita Schmitz-Koep, Dr.
Department of Diagnostic and Interventional Neuroradiology, School of Medicine
Munich, Bavaria
Department of Diagnostic and Interventional Neuroradiology, School of Medicine
Munich, Bavaria
Kirsten Jung
Department of Diagnostic and Interventional Neuroradiology, School of Medicine
Munich, Bavaria
Department of Diagnostic and Interventional Neuroradiology, School of Medicine
Munich, Bavaria
Dennis Hedderich
Department of Diagnostic and Interventional Neuroradiology, School of Medicine
Munich, Bavaria
Department of Diagnostic and Interventional Neuroradiology, School of Medicine
Munich, Bavaria
Introduction:
Structural magnetic resonance imaging (MRI) continues to inhabit an important position in research and clinical assessments of neurodegeneration, with one of its primary uses lying in the detection and monitoring of brain atrophy patterns (Knopman et al., 2016; Young et al., 2020). In clinical practice however, structural imaging protocols are often heterogeneous and subject to low spatial resolution, resulting in suboptimal image quality for brain volume assessments and limited data quality for scientific analyses (Iglesias et al., 2023).
Addressing this issue, Freesurfer is a widely used open-source software package, employed among other use-cases in volumetric and surface-based analyses of neuroimaging (Reuter, Schmansky, Rosas, & Fischl, 2012). Since the recent release of version 7.3, Freesurfer includes SynthSR, a convolutional neural network based approach able to generate 1 mm isotropic 3D T1-like synthetic imaging (T1s) from heterogeneous input sequences trained on data from 20 subjects (Iglesias et al., 2023). In previous validation approaches, the developers report strong correlations between T1s and real 3D T1 imaging (Iglesias et al., 2023).
Reliable generation of T1s could improve clinical brain atrophy assessments and unlock much larger datasets of neurodegeneration-related imaging than currently available. In the present study, we attempt to further validate T1s against the gold standard of 1 mm isotropic 3D T1 imaging (GS) by investigating bilateral hippocampus volume (VHip), a notable imaging parameter in neurodegeneration assessment in patients and healthy controls (HC) (Knopman et al., 2016).
Addressing this issue, Freesurfer is a widely used open-source software package, employed among other use-cases in volumetric and surface-based analyses of neuroimaging (Reuter, Schmansky, Rosas, & Fischl, 2012). Since the recent release of version 7.3, Freesurfer includes SynthSR, a convolutional neural network based approach able to generate 1 mm isotropic 3D T1-like synthetic imaging (T1s) from heterogeneous input sequences trained on data from 20 subjects (Iglesias et al., 2023). In previous validation approaches, the developers report strong correlations between T1s and real 3D T1 imaging (Iglesias et al., 2023).
Reliable generation of T1s could improve clinical brain atrophy assessments and unlock much larger datasets of neurodegeneration-related imaging than currently available. In the present study, we attempt to further validate T1s against the gold standard of 1 mm isotropic 3D T1 imaging (GS) by investigating bilateral hippocampus volume (VHip), a notable imaging parameter in neurodegeneration assessment in patients and healthy controls (HC) (Knopman et al., 2016).
Methods:
We selected a dataset of 10 representative Alzheimer's Disease (AD) cases, as well as 10 HC scanned on a 3T Siemens Biograph scanner in our local clinic. We employed Freesurfer SynthSR (Iglesias et al., 2023) to generate T1s from three different scenarios of imaging input: 1 mm 3D isotropic T2 FLAIR (Sc1), 4 mm axial T2 FLAIR (Sc2) and 4 mm coronal T2 (Sc3). The resulting T1s and GS were further segmented via CAT12 according to the LONI Probabilistic Brain Atlas (LPBA40) (Gaser et al., 2022; Shattuck et al., 2008). VHip were extracted from GS and the three sets of T1s for subsequent testing against one another via paired t-tests.
Results:
After Bonferroni correction, we observed significantly higher VHip in T1s based on Sc3 compared to GS in AD (Figure 1; GS 5.867 ± 0.580 ml vs. Sc3 6.368 ± 0.604 ml, p = 0.01126). Notably, no significant differences were observed between GS and T1s of HC (GS 7.183 ± 0.644 ml; Sc1 6.979 ± 0.889 ml; Sc2 6.769 ± 0.742 ml; Sc3 7.061 ± 0.750 ml). T1s VHip overestimation was strong enough in some cases to be visually notable (Figure 2).
Conclusions:
Overall, the performance of SynthSR was not significantly different from GS for any input scenarios aside from Sc3. Regarding potential scientific and clinical use cases for T1s, these are generally encouraging results confirming high congruence with GS.
Nonetheless, we observed significant overestimation of VHip in T1s synthesized from Sc3 in our limited sample of AD. One potential reason for this could lie within the training data employed in the generation of SynthSR, which despite including imaging from probable Alzheimer's cases (Iglesias et al., 2023) could introduce biases towards healthy brain volumina. This could consequently facilitate potentially faulty interpolations for e. g. partial volume effects (Figure 2). Future studies should consider validation of T1s in atypical patterns of atrophy, such as e. g. frontotemporal lobar degeneration.
Nonetheless, we observed significant overestimation of VHip in T1s synthesized from Sc3 in our limited sample of AD. One potential reason for this could lie within the training data employed in the generation of SynthSR, which despite including imaging from probable Alzheimer's cases (Iglesias et al., 2023) could introduce biases towards healthy brain volumina. This could consequently facilitate potentially faulty interpolations for e. g. partial volume effects (Figure 2). Future studies should consider validation of T1s in atypical patterns of atrophy, such as e. g. frontotemporal lobar degeneration.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Lifespan Development:
Aging
Modeling and Analysis Methods:
Image Registration and Computational Anatomy 2
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Cortical Anatomy and Brain Mapping
Novel Imaging Acquisition Methods:
Anatomical MRI
Keywords:
Aging
Open-Source Software
Segmentation
STRUCTURAL MRI
Other - synthetic MRI, validation, atrophy
1|2Indicates the priority used for review
Provide references using author date format
Iglesias, J. E., Billot, B., Balbastre, Y., Magdamo, C., Arnold, S. E., Das, S., . . . Fischl, B. (2023). SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science advances, 9(5), eadd3607.
Knopman, D. S., Jack, C. R., Wiste, H. J., Weigand, S. D., Vemuri, P., Lowe, V. J., . . . Mielke, M. M. (2016). Age and neurodegeneration imaging biomarkers in persons with Alzheimer disease dementia. Neurology, 87(7), 691-698.
Reuter, M., Schmansky, N. J., Rosas, H. D., & Fischl, B. (2012). Within-subject template estimation for unbiased longitudinal image analysis. Neuroimage, 61(4), 1402-1418.
Shattuck, D. W., Mirza, M., Adisetiyo, V., Hojatkashani, C., Salamon, G., Narr, K. L., . . . Toga, A. W. (2008). Construction of a 3D probabilistic atlas of human cortical structures. Neuroimage, 39(3), 1064-1080.
Young, P. N., Estarellas, M., Coomans, E., Srikrishna, M., Beaumont, H., Maass, A., . . . Betts, M. J. (2020). Imaging biomarkers in neurodegeneration: current and future practices. Alzheimer's Research & Therapy, 12(1), 1-17.