123
Poster No:
2010
Submission Type:
Abstract Submission
Authors:
Fraser Aitken1, Donald Tournier2, Joseph Hajnal2, Paul Cawley3, David Carmichael4
Institutions:
1Kings Collge London, London, London, 2King's College London, London, England, 3King's College London, London, United Kingdom, 4King's College London, London, Please select an option below
First Author:
Co-Author(s):
Introduction:
Ultra-low/low field MRI has drawn renewed neuroscientific interest due to lower cost and portability that may increase access (Hori et al., 2021). Several possible clinical applications require accurate head tissue segmentation (Michel & Brunet, 2019). Images acquired at lower field strengths have reduced SNR, image resolution and may also suffer from image distortions compared to higher field/cost scanners (Arnold et al., 2023). The contribution of these factors on segmentation reliability are unclear. We aimed to quantify the accuracy of automated head tissue segmentations using CHARM (Thielscher et al., 2015) on 0.064T and 0.55T images in comparison to those from a 3T image.
Methods:
Data acquisition
3D T1 and T2 weighted (w) images were acquired in 2 healthy adults (ages 35 and 41) at 0.064 T (Hyperfine Swoop™), 0.55T (Siemens Healthineers MAGNETOM, Free.Max) and 3 T ((Siemens Healthineers MAGNETOM, Vida).
T1w parameters. Hyperfine, inversion recovery FSE 2x2x2 mm, TR=880 ms, TE=6.12 ms. TI=354. Freemax, 3D FLASH, TR =17 ms, TE=10.70 ms, 3T Vida, MPRAGE, 0.9 x 0.9 x0.9 mm, TR=2200 TE=2.46 ms. T2w parameters. Hyperfine 2x2x2 mm, TR=2000ms, TE=156 ms. Freemax T2 SPACE, TR=1300 ms, TE= 197 ms. Vida T2 SPACE, TR=3200ms, TE=408ms.
Preprocessing
Linear co-registration
T1 and T2 images from all 3 scanners were oriented to the same reference location using SPM 12. Linear co-registrations were carried out using SPM 12(Penny et al., 2007). T1 and T2 images from all field strengths were co-registered and resliced to the 3T T1w image.
Segmentation creation.
Head models were created using CHARM, a part of SimNIBS. CHARM requires a T1-weighted image and preferably a T2-weighted image to create a segmentation of the head into several tissue types, including white matter, grey matter, CSF, bone, scalp, eyeballs, compact bone, spongy bone, blood, and muscle.
Nonlinear co-registration of segmented images
To correct for distortion the following nonlinear registration was performed: (i) the skull segment was extracted from all segmentations as a binary mask; (ii) rigid & nonlinear registration was performed using mrregister (a part of MRtrix3 (Tournier et al., 2019)) to align the 0.064T & 0.55T skull segments with the 3T skull segment, using a least-squares metric with strong smoothness regularization as the geometric distortions were expected to be smooth (e.g. due to B0 and gradient non-linearity); (iii) the resulting warps were applied to the full 0.064T & 0.55T segmentations using mrtransform (also a part of MRtrix 3) using nearest-neighbor interpolation.
Data analysis
Low-field segmentation quality was assessed by comparison with the reference 3T segmentation, quantified using the Dice coefficient
3D T1 and T2 weighted (w) images were acquired in 2 healthy adults (ages 35 and 41) at 0.064 T (Hyperfine Swoop™), 0.55T (Siemens Healthineers MAGNETOM, Free.Max) and 3 T ((Siemens Healthineers MAGNETOM, Vida).
T1w parameters. Hyperfine, inversion recovery FSE 2x2x2 mm, TR=880 ms, TE=6.12 ms. TI=354. Freemax, 3D FLASH, TR =17 ms, TE=10.70 ms, 3T Vida, MPRAGE, 0.9 x 0.9 x0.9 mm, TR=2200 TE=2.46 ms. T2w parameters. Hyperfine 2x2x2 mm, TR=2000ms, TE=156 ms. Freemax T2 SPACE, TR=1300 ms, TE= 197 ms. Vida T2 SPACE, TR=3200ms, TE=408ms.
Preprocessing
Linear co-registration
T1 and T2 images from all 3 scanners were oriented to the same reference location using SPM 12. Linear co-registrations were carried out using SPM 12(Penny et al., 2007). T1 and T2 images from all field strengths were co-registered and resliced to the 3T T1w image.
Segmentation creation.
Head models were created using CHARM, a part of SimNIBS. CHARM requires a T1-weighted image and preferably a T2-weighted image to create a segmentation of the head into several tissue types, including white matter, grey matter, CSF, bone, scalp, eyeballs, compact bone, spongy bone, blood, and muscle.
Nonlinear co-registration of segmented images
To correct for distortion the following nonlinear registration was performed: (i) the skull segment was extracted from all segmentations as a binary mask; (ii) rigid & nonlinear registration was performed using mrregister (a part of MRtrix3 (Tournier et al., 2019)) to align the 0.064T & 0.55T skull segments with the 3T skull segment, using a least-squares metric with strong smoothness regularization as the geometric distortions were expected to be smooth (e.g. due to B0 and gradient non-linearity); (iii) the resulting warps were applied to the full 0.064T & 0.55T segmentations using mrtransform (also a part of MRtrix 3) using nearest-neighbor interpolation.
Data analysis
Low-field segmentation quality was assessed by comparison with the reference 3T segmentation, quantified using the Dice coefficient
Results:
Dice co-efficients (see figure 1) indicate that segmentation quality based on 0.55T data were high (~0.7-0.95) indicating a similar segmentation to that from 3T was achieved. . In contrast, dice co-efficients for segmentations derived from the 0.064T images were significantly lower (~0.25-0.5). This appeared to be related to geometric image distortions (see figure 2). Crucially, with a nonlinear registration of the segmented images there was a marked improvement in dice coefficients for the 0.064T data (~0.65-0.82).
Conclusions:
Dice coefficients based on our segmented images based on 0.55T were high, indicating a very good level of similarity with our 3T derived data. On the other hand dice coefficients derived from 0.064T data were relatively low. The segmentation similarity was substantially improved by nonlinear registration. This indicated that at 0.064T image noise and contrast might be sufficient, but improved distortion (without requiring matched high-field data) is needed.
Modeling and Analysis Methods:
Methods Development 2
Segmentation and Parcellation 1
Novel Imaging Acquisition Methods:
Anatomical MRI
Keywords:
MRI
Segmentation
1|2Indicates the priority used for review
·Figure 1
·Figure 2
Provide references using author date format
Hori, M., Hagiwara, A., Goto, M., Wada, A., & Aoki, S. (2021). Low-Field Magnetic Resonance Imaging: Its History and Renaissance. Investigative Radiology, 56(11), 669. https://doi.org/10.1097/RLI.0000000000000810
Michel, C. M., & Brunet, D. (2019). EEG source imaging: A practical review of the analysis steps. Frontiers in Neurology, 10(APR), 446653. https://doi.org/10.3389/FNEUR.2019.00325/BIBTEX
Penny, W., Friston, K., Ashburner, J., Kiebel, S., & Nichols, T. (2007). Statistical Parametric Mapping: The Analysis of Functional Brain Images. Statistical Parametric Mapping: The Analysis of Functional Brain Images. https://doi.org/10.1016/B978-0-12-372560-8.X5000-1
Thielscher, A., Antunes, A., & Saturnino, G. B. (2015). Field modeling for transcranial magnetic stimulation: A useful tool to understand the physiological effects of TMS? Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS, 2015-November, 222–225. https://doi.org/10.1109/EMBC.2015.7318340
Tournier, J. D., Smith, R., Raffelt, D., Tabbara, R., Dhollander, T., Pietsch, M., Christiaens, D., Jeurissen, B., Yeh, C. H., & Connelly, A. (2019). MRtrix3: A fast, flexible and open software framework for medical image processing and visualisation. NeuroImage, 202. https://doi.org/10.1016/J.NEUROIMAGE.2019.116137