pubmed.ncbi.nlm.nih.gov

Fast cortical surface reconstruction from MRI using deep learning - PubMed

  • ️Sat Jan 01 2022

Fast cortical surface reconstruction from MRI using deep learning

Jianxun Ren et al. Brain Inform. 2022.

Abstract

Reconstructing cortical surfaces from structural magnetic resonance imaging (MRI) is a prerequisite for surface-based functional and anatomical image analyses. Conventional algorithms for cortical surface reconstruction are computationally inefficient and typically take several hours for each subject, causing a bottleneck in applications when a fast turnaround time is needed. To address this challenge, we propose a fast cortical surface reconstruction (FastCSR) pipeline by leveraging deep machine learning. We trained our model to learn an implicit representation of the cortical surface in volumetric space, termed the "level set representation". A fast volumetric topology correction method and a topology-preserving surface mesh extraction procedure were employed to reconstruct the cortical surface based on the level set representation. Using 1-mm isotropic T1-weighted images, the FastCSR pipeline was able to reconstruct a subject's cortical surfaces within 5 min with comparable surface quality, which is approximately 47 times faster than the traditional FreeSurfer pipeline. The advantage of FastCSR becomes even more apparent when processing high-resolution images. Importantly, the model demonstrated good generalizability in previously unseen data and showed high test-retest reliability in cortical morphometrics and anatomical parcellations. Finally, FastCSR was robust to images with compromised quality or with distortions caused by lesions. This fast and robust pipeline for cortical surface reconstruction may facilitate large-scale neuroimaging studies and has potential in clinical applications wherein brain images may be compromised.

Keywords: Cortical surface reconstruction; Deep learning; FreeSurfer; Level set; T1-weighted MRI.

© 2022. The Author(s).

PubMed Disclaimer

Conflict of interest statement

H.L. is on the chief scientific advisory board for Neural Galaxy LLC. The other authors declare no competing financial interests.

Figures

Fig. 1
Fig. 1

Schematic of the fast cortical surface reconstruction (FastCSR) workflow. The FastCSR workflow can be summarized in four steps: a original T1w images are normalized and fed to a 3D U-Net for segmentation of white matter from gray matter. b After segmentation, hemispheric white matter masks are generated, distinguishing the two hemispheres in the T1w images. c The hemispheric masks and the original T1w images are fed to another 3D U-Net to predict the level set representation of the cortical surface. Level set is an implicit representation of the cortical surface. The U-shaped network architecture can be briefly described as follows: each blue arrow represents a 3 × 3 × 3 convolution process followed by a leaky rectified linear unit (LReLU). Each orange downward arrow indicates a 2 × 2 × 2 max pooling. Each red upward arrow indicates a 2 × 2 × 2 upsampling followed by a convolution (Up-Conv). The long-range skip connections using the copy-and-concatenate (Copy & Concat) operation are indicated by green arrows. d The level set representations of the surfaces are generated by the deep learning model. The voxels whose level set value equals to zero delineate the boundary of the cortical surface. Negative voxels indicated by dark colors are below the surface and positive voxels indicated by light colors are above the surface. The red box shows the level set representation in the left frontal cortex magnified to better visualize the details in the surface boundary. e An explicit surface mesh is reconstructed from the level set representation through a fast topology-preserving isosurface extraction algorithm. The resulting surface is visualized using a dorsal orientation

Fig. 2
Fig. 2

Processing time of FastCSR, FreeSurfer and FastSurfer. The bar graphs illustrate differences in the average processing time for the validation dataset using FreeSurfer, FastSurfer and FastCSR. We compared the processing times for the surface reconstruction step alone (surface recon) and the whole pipeline including all necessary preprocessing steps for surface reconstruction using either sequential or parallel processing, the latter of which processes both hemispheres simultaneously. Among the three methods, FastCSR achieved the highest computational efficiency for the surface reconstruction step (5.22 ± 0.92 min and 2.61 ± 0.46 min for sequential and parallel processing, respectively), as well as for the whole pipeline (7.05 ± 0.92 min and 4.44 ± 0.46 min for sequential and parallel processing, respectively). Error bars indicate standard deviations

Fig. 3
Fig. 3

Surfaces reconstructed by FastCSR are comparable to results from FreeSurfer. a The cortical surfaces from a randomly selected healthy participant reconstructed using the FreeSurfer (the upper panel) and the FastCSR (the lower panel) pipelines. The surfaces generated by different methods show high similarity with small discrepancies that are highlighted by orange dashed boxes. b Horizontal slices at multiple levels taken from the participant’s T1w image showing the cortical surface identified using FreeSurfer (yellow lines) and FastCSR (red lines) show that both methods accurately capture the boundary between white and gray matter. The surfaces derived from these two methods showed a high degree of overlap in most of cortical areas, indicating high concordance between these two methods. Small discrepancies are highlighted by orange dashed boxes

Fig. 4
Fig. 4

Differences in surfaces reconstructed by FastCSR, FreeSurfer and FastSurfer. a To quantitatively assess the surface agreement, we measured the average surface displacement between FastCSR and FreeSurfer across participants in the validation set. The maximal displacement between FreeSurfer and FastCSR was smaller than 0.5 mm, which is approximately half of the voxel size. b The average surface displacement between FastSurfer and FreeSurfer was also measured. The average surface displacement map showed a similar pattern with that observed from using the FastCSR method (Spearman’s ρ = 0.714, p < 0.0001). c The direct comparison of displacement maps between the FastCSR vs. FreeSurfer contrast and the FastSurfer vs. FreeSurfer contrast was performed. Results showed that the FastCSR approach achieved overall better performance. Lateral gyri and visual cortices showed significantly smaller displacement in the FastCSR versus FreeSurfer method than FastSurfer versus FreeSurfer (two-tailed paired t-tests, p < 0.01, FDR corrected)

Fig. 5
Fig. 5

Surfaces reconstructed by FastCSR and FreeSurfer show comparable mesh quality. To assess the mesh quality of the FastCSR surface, we estimated the mesh quality in terms of Q value in the validation set. The Q values of FastCSR (blue dots) and FreeSurfer (purple dots) are illustrated using swarm plots, with the means of each distribution depicted by boxplots with boxes marks the high and low quartiles and whiskers indicating the minimum and maximum values. The Q values obtained with FastCSR (QFastCSR = 0.903 ± 0.002) are significantly higher than those of FreeSurfer (QFreeSurfer = 0.886 ± 0.003; two-tailed paired t-test, t(161) = 122.008, p = 1.462 × 10–160), indicating greater mesh quality achieved with our approach

Fig. 6
Fig. 6

Surface morphometries measured in unseen datasets. To examine if FastCSR is generalizable to unseen datasets, we applied this method to the previously unseen ABIDE-II dataset with T1w images at 1.0-mm resolution and the HCP dataset with 0.7-mm resolution images. These data were also processed using the FreeSurfer pipeline. Cortical thickness and sulcal depth were estimated. a The average cortical thickness maps obtained from FreeSurfer (left) versus FastCSR (right) in the ABIDE-II dataset are similar, with only 0.07% of the vertices demonstrating significant difference (two-tailed paired t-tests, p < 0.01, FDR corrected). b For the HCP dataset, the average cortical thickness maps derived from FreeSurfer and FastCSR are also similar, with only 2.16% of the vertices showing significant difference (two-tailed paired t-tests, p < 0.01, i.e., − log10(p) > 2.0, FDR corrected). c The positive values in the sulcal depth maps indicate sulci (warm colors) and negative values indicate gyri (cool colors). For the ABIDE-II dataset, 2.44% of the vertices showed significant differences between the FreeSurfer and FastCSR method. Differences were mainly distributed in the insular cortices, the precentral gyrus, and the medial orbitofrontal cortices. d For the HCP dataset, 5.87% of the vertices showed significant difference in sulcal depth between FreeSurfer and FastCSR

Fig. 7
Fig. 7

Anatomical parcellation in the ASD and HCP datasets. We assessed the similarity in anatomical cortical parcellation, measured by the Dice coefficient, for each cortical region obtained with FastCSR compared to FreeSurfer in both the ABIDE-II (left) and HCP (right) datasets. The Dice coefficients for most cortical areas (77.94% in ABIDE and 76.47% in HCP) are above 95%. Cortical areas with Dice coefficients smaller than 90% included the entorhinal region and the left rostral anterior cingulate cortex in both datasets, and the in the HCP dataset, additionally included the left parahippocampal area

Fig. 8
Fig. 8

Surface morphometries and anatomical parcellation from FastCSR showed high intra-subject test–retest reliability. To examine the reliability of our FastCSR method, we measured the instability of surface morphometries and anatomical parcellation in a dataset consisting of 30 participants with 10 repeated scans for each participant. A The instability of morphometrics and parcellations was estimated by the standard deviation across the 10 sessions in each vertex for each participant. The lower instability, indicated by red color, suggests higher test–retest reliability. The average instability map across 30 individuals showed similar distributions for both the FreeSurfer (the upper panel) and FastCSR (the lower panel) methods. However, the FastCSR show lower instability for cortical thickness, sulcal depth, and parcellation than FreeSurfer. B Histograms illustrate the distribution of measurements obtained from FastCSR (blue bars) and FreeSurfer (purple bars). FastCSR shows lower instability relative to FreeSurfer in measures of cortical thickness (two-sample Kolmogorov–Smirnov test, p = 1.130 × 10–7), sulcal depth (two-sample Kolmogorov–Smirnov test, p = 9.700 × 10–3), and anatomical parcellation (two-sample Kolmogorov–Smirnov test, p = 1.037 × 10–37)

Fig. 9
Fig. 9

FastCSR is robust against image quality and brain distortions. a Cortical surfaces of two individuals with poor imaging quality from the CoRR-HNU dataset are reconstructed by FreeSurfer and FastCSR. White arrows highlight the jagged gyri caused by noise in the image (see upper panel) using the FreeSurfer pipeline. Surfaces obtained from the same individuals are reconstructed using our FastCSR method (see lower panel) and yielded cortical surfaces with smoother gyri. b FreeSurfer failed to reconstruct brain surfaces for three stroke patients whose brains are distorted due to lesions, whereas our FastCSR successfully reconstructed the cortical surfaces in these patients. The anatomical boundaries demarcating white matter and pial boundaries (yellow lines) overlaid onto horizontal sections of the T1w images with white arrows indicating the stroke lesions (upper panel). The corresponding cortical surfaces of the lesional hemispheres are displayed (lower panel). The white arrows indicate the stroke lesions

Similar articles

Cited by

References

    1. Dale AM, Fischl B, Sereno MI. Cortical surface-based analysis. I. Segmentation and surface reconstruction. Neuroimage. 1999;9:179–194. - PubMed
    1. Griffin LD. The intrinsic geometry of the cerebral cortex. J Theor Biol. 1994;166:261–273. - PubMed
    1. Han X, Pham DL, Tosun D, et al. CRUISE: cortical reconstruction using implicit surface evolution. Neuroimage. 2004;23:997–1012. - PubMed
    1. Fischl B, Dale AM. Measuring the thickness of the human cerebral cortex from magnetic resonance images. Proc Natl Acad Sci U S A. 2000;97:11050–11055. - PMC - PubMed
    1. Kruggel F, Bruckner MK, Arendt T, et al. Analyzing the neocortical fine-structure. Med Image Anal. 2003;7:251–264. - PubMed

LinkOut - more resources