Recon-any: A FreeSurfer-based cortical reconstruction tool for any MRI
Poster No:
1934
Submission Type:
Abstract Submission
Authors:
Karthik Gopinath1, Oula Puonti1,2, Douglas Greve1, Annabel Sorby-Adams3, Jennifer Guo3, Sudeshna Das3, Steve Arnold3, Colin Magdamo3, Mark Montine4, Caitlin Latimer4, Christine Mac Donald4, C. Dirk Keene4, W. Taylor Kimberly3, Juan Eugenio Iglesias1
Institutions:
1Athinoula A. Martinos Center for Biomedical Imaging, MGH, HMS, Boston, MA, 2Danish Research Insttitute for Magnetic Resonance, Hvidovre, Region H, 3Massachusetts General Hospital, Boston, MA, 4University of Washington, Seattle, WA
First Author:
Co-Author(s):
Oula Puonti
Athinoula A. Martinos Center for Biomedical Imaging, MGH, HMS|Danish Research Insttitute for Magnetic Resonance
Boston, MA|Hvidovre, Region H
Athinoula A. Martinos Center for Biomedical Imaging, MGH, HMS|Danish Research Insttitute for Magnetic Resonance
Boston, MA|Hvidovre, Region H
Introduction:
Human brain MRI scanning protocols in research almost always include a 1mm isotropic T1 scan for morphometric analysis, and most image analysis pipelines are designed for this specific resolution and MRI contrast. However, there are other imaging modalities that are incompatible with most current tools. Ex vivo MRI is useful for high-resolution mesoscopic studies but does not have T1 contrast[1]. Portable MRI enables affordable large-scale scanning but has low resolution and signal-to-noise ratio[2]. Existing clinical MRI scans are abundant in hospitals and can provide vast amounts of data "for free" but often have large slice spacing and heterogeneous MRI contrast. More recently, 3D-reconstructed dissection photographs enable postmortem morphometry[3] but have large-spacing and heterogeneous image contrast. Here we propose "Recon-any": a deep learning enabled version of FreeSurfer's[4,5] "recon-all" that can handle imaging volumes of any resolution and contrast "out of the box" without retraining.
Methods:
Recon-any has two main parts: 1) estimation of isotropic Signed Distance Functions (SDFs) from anisotropic imaging volumes (MRI scans or 3D reconstructed photographs) and 2) placing the white matter (WM) and pial surfaces and subsequent cortical analysis.
Learning of SDFs: The learning module simulates synthetic imaging volumes from 3D segmentations, with random orientation, resolution, and contrast. It uses a Gaussian mixture model[6,7] and random bias field, resolution, and noise simulators. The regression targets are voxel-wise SDFs precomputed from WM and pial meshes. A regression convolutional neural network is trained by optimizing weights to minimize the L1 norm of the difference between predicted and ground truth SDFs. At test time, the trained network predicts 1mm isotropic SDFs for any imaging volume.
Geometry Processing for Surface Placement: For WM surface reconstruction, we first fill in holes in the WM mask and tessellate it to obtain the initial mesh. Then, we smooth the mesh and use automated manifold surgery[5] to guarantee spherical topology. Next, we iteratively deform the WM mesh by minimizing an objective function consisting of a fidelity term (based on the absolute value of the SDF) and a regularizer. The pial surface is fitted using the predicted SDF of the pial surface. The resulting surfaces undergo FreeSurfer analysis for cortical thickness, parcellation, and registration to an atlas in spherical coordinates.
Learning of SDFs: The learning module simulates synthetic imaging volumes from 3D segmentations, with random orientation, resolution, and contrast. It uses a Gaussian mixture model[6,7] and random bias field, resolution, and noise simulators. The regression targets are voxel-wise SDFs precomputed from WM and pial meshes. A regression convolutional neural network is trained by optimizing weights to minimize the L1 norm of the difference between predicted and ground truth SDFs. At test time, the trained network predicts 1mm isotropic SDFs for any imaging volume.
Geometry Processing for Surface Placement: For WM surface reconstruction, we first fill in holes in the WM mask and tessellate it to obtain the initial mesh. Then, we smooth the mesh and use automated manifold surgery[5] to guarantee spherical topology. Next, we iteratively deform the WM mesh by minimizing an objective function consisting of a fidelity term (based on the absolute value of the SDF) and a regularizer. The pial surface is fitted using the predicted SDF of the pial surface. The resulting surfaces undergo FreeSurfer analysis for cortical thickness, parcellation, and registration to an atlas in spherical coordinates.
Results:
We evaluate our method's performance on two datasets, HCP[8] (15 subjects, with T1 and T2 scans) and ADNI3[9] (15 subjects, with T1 and FLAIR scans). To evaluate performance as a function of resolution and plane of acquisition, we downsample the initial isotropic data to 2, 4, and 6mm resolution in axial, coronal, and sagittal direction. Subsequently, we quantify the disparity between the regional cortical thicknesses estimated by Recon-any and the ground truth, which is the output of FreeSurfer's recon-all on the 1mm T1 scans. Recon-any achieves a mean absolute error of 0.28mm, 0.26mm, and 0.29mm at 2mm, 4mm, and 6mm for FLAIR scans, and 0.21mm, 0.26mm, and 0.27mm for T2 scans. Additionally, the accuracy of the Desikan-Killiany parcellation[10] of our method against the ground truth reveals Dice scores consistently above 0.91 for every resolution, with an average Dice score of 0.95. We present qualitative results for T2w, ex vivo, 3D photo volume, and portable MRI in Figure 1. Quantitative assessments with paired 1mm T1 data remain as future work.
Conclusions:
We have presented a cortical analysis method that works "out of the box" for any MRI contrast and resolution. Future work will seek to replace geometry processing modules with faster learning methods and improve the reliability of thickness measurements. We will make our method publicly available on FreeSurfer, enabling researchers worldwide to perform cortical analysis on large amounts of data which is currently not possible.
Modeling and Analysis Methods:
Image Registration and Computational Anatomy
Methods Development 1
Segmentation and Parcellation 2
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Cortical Anatomy and Brain Mapping
Keywords:
Cortex
Machine Learning
MRI
1|2Indicates the priority used for review
Provide references using author date format
[2] Taylor, K.W., Sorby-Adams, A.J., Webb, A.G., Wu, E.X., Beekman, R., Bowry, R., Schiff, S.J., et al. (2023), 'Brain imaging with portable low-field MRI', Nature Reviews Bioengineering, vol. 1, no. 9, pp. 617-630.
[3] Gazula, H., Tregidgo, H.F., Billot, B., Balbastre, Y., William-Ramirez, J., Herisse, R., Deden-Binder, L.J., Casamitjana, A., Melief, E.J., Latimer, C.S., and Kilgore, M.D. (2023), 'Machine learning of dissection photographs and surface scanning for quantitative 3D neuropathology', bioRxiv, pp.2023-06.
[4] Dale, A.M., Fischl, B., and Sereno, M.I. (1999), 'Cortical surface-based analysis: I. segmentation and surface reconstruction', Neuroimage, vol. 9, pp. 179–19.
[5] Fischl, B., Sereno, M., and Dale, A.M. (1999), 'Cortical surface-based analysis: II: inflation, flattening, and a surface-based coordinate system', Neuroimage, vol. 9, pp. 195–207.
[6] Billot, B., Greve, D.N., Puonti, O., Thielscher, A., Van Leemput, K., Fischl, B., Dalca, A.V., Iglesias, J.E., et al. (2023a), 'Synthseg: Segmentation of brain MRI scans of any contrast and resolution without retraining', Medical Image Analysis, vol. 86, 102789.
[7] Iglesias, J., Billot, B., Balbastre, Y., Tabari, A., et al. (2021), 'Joint super-resolution and synthesis of 1 mm isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution & contrast', Neuroimage, vol. 237, 118206.
[8] Glasser, M., Sotiropoulos, S., Wilson, J.A., and Coalson, T.S. (2013), 'The minimal preprocessing pipelines for the Human Connectome Project', Neuroimage, vol. 80, pp. 105–24.
[9] Jack Jr, C.R., Bernstein, M.A., Fox, N.C., Thompson, P., Alexander, G., et al. (2008), 'The Alzheimer’s disease neuroimaging initiative (ADNI)', Journal of Magnetic Resonance Imaging, vol. 27, pp. 685–691.
[10] Desikan, R.S., S ́egonne, F., Fischl, B., Quinn, B.T., Dickerson, B.C., Blacker, D., Buckner, R.L., Dale, A.M., et al. (2006), 'An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest', Neuroimage, vol. 31, pp. 968–980.