2159

Infant Brain Segmentation based on Multi-dMRI contrast Attention (BISMA)

Tianjia Zhu^1,2, Minhui Ouyang¹, Lei Feng³, Kay Sindabizera¹, Jessica Hyland¹, and Hao Huang^1,4
¹Department of Radiology, Children's Hospital of Philadelphia, Philadelphia, PA, United States, ²Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, United States, ³Shandong University Cheeloo College of Medicine, Jinan, China, ⁴Department of Radiology, University of Pennsylvania, Philadelphia, PA, United States

Synopsis

On widely used relaxation-based T1-weighted (T1w) and T2-weighted (T2w) MRI images, 0-2-year-old infant brain images exhibit poor gray and white matter contrasts due to dynamic and poor myelination in this stage. Such poor T1w and T2w contrasts hinder accurate segmentation of infant brains based on T1w or T2w images. Instead, diffusion MRI (dMRI)-derived maps offer high contrasts throughout infancy. To leverage rich dMRI contrasts, we established an Infant Brain Segmentation based on Multi-dMRI contrast Attention (BISMA) technique. BISMA fills the gap in accurate deep learning segmentation of infant brain by fusing multi-contrasts of dMRI-derived maps.

Purpose

On most widely used traditional relaxation-based T1-weighted (T1w) and T2-weighted (T2w) MRI images, 0-2-year-old infant brain images exhibit large within-tissue intensity variations, and regionally heterogenous and dynamic contrast changes due to the ongoing myelination process in infant brains¹, resulting in poor contrast among brain gray matter (GM), white matter (WM), and cerebral spinal fluid. Existing segmentation algorithms relying on relaxation-based contrasts consequently achieve relatively low accuracy² (Fig. 1). Diffusion MRI (dMRI)-derived maps offer high inter-tissue contrasts and high intra-tissue consistency throughout infant brain development due to their robustness to the myelination process^3-5. In this study, we establish an Infant Brain Segmentation based on Multi-dMRI contrast Attention (BISMA) technique to leverage the rich information from dMRI contrasts. BISMA fills the gap in deep learning (DL) segmentation of infant brain fusing multi-contrasts of dMRI maps.

Methods

Acquisition of structural and diffusion MRI: With approval from the institutional review board, a cohort of 89 neonates with dMRI and T2w images was acquired. DMRI was at 1.5×1.5×1.6mm³ resolution with b=1000 and 1600 s/mm², 30 gradient directions for each b-value. T2w was acquired with the same resolution. Diffusion tensor imaging (DTI) fitting: After correction of eddy current distortion and affine registration of diffusion weighted images, DTI metrics were fitted. Fractional anisotropy (FA), axial (AD), radial (RD) and mean diffusivity (MD) maps were obtained (Fig. 2A-2B). Manual segmentation for ground-truth: Manual segmentation of the datasets from 14 neonates based on DTI-derived maps was done by an experienced neuroanatomist (Fig. 2A, left). Segmentation on single contrast: A UNet model⁶ was trained to segment neonate brains using single dMRI contrast, with 13 subjects used for training and 1 subject leaved for validation. The network depth was 5. Starting learning rate was 0.01. StepLR scheduler with momentum=0.9 and stochastic gradient descent (SDG) optimizer was used. The network was trained for 10 epochs. Segmentation by BISMA with AD and RD contrasts: We adopted the convolutional bottleneck attention module (CBAM)⁷ into a UNet architecture (Fig. 3) to form the BISMA model. At each level of the UNet, CBAM is applied to the extracted feature maps. The BISMA model was trained using the same parameters as UNet for single contrast.

Results

For 3-6-month-old infants, segmentation based on T1w and T2w MRI usually suffers from low accuracy due to the poor contrasts (Fig. 1A-1B). Yellow contour on enlarged T1w (Fig. 1D), and T2w (Fig. 1E) show the ground-truth segmentation, whereas yellow contour in Fig. 1F shows the T1w-based segmentation done by Infant Freesurfer² deviating from the ground-truth. DTI contrasts offer rich complementary information about the underlying tissue types (Fig. 2). Fig. 2A shows the ground-truth segmentation along with b0 and FA contrasts on a representative neonate subject. RD and MD offer superior contrast for cortical GM (green arrows), where as AD offers the best contrast for deep GM (blue arrows). FA outperforms other contrasts for showing myelinated WM (yellow arrows). UNet⁶ segmentations of GM (Fig. 4A) and WM (Fig. 4B) on single contrast along with segmentation from BISMA combining AD and RD contrasts exhibit differential regional accuracies (Fig. 4A-4B) compared with ground-truth (Fig. 4A-B right-most column). Enlarged cortical GM segmentation (Fig. 4C) demonstrate superior performance of BISMA in comparison to that from any single contrast. Enlarged deep GM (Fig. 4D) also demonstrates superior performance of BIMSA in segmenting deep GM, where AD alone under-estimates and RD alone overestimates these areas.

Similarly, Fig. 4E demonstrates the performance of BISMA in delineating internal capsule in WM in comparison to single contrasts. BISMA with AD+RD as inputs improves overall GM Dice by 6% (Fig. 4F) to 0.93. WM BISMA Dice coefficient reaches 0.89, improved by >10% compared to single contrasts (Fig. 4G). All Dice improvements are larger than common inter-rater segmentation error of 5% for neonate brain MRI⁸.

Discussion and Conclusion

Our results show single DTI contrasts’ sensitivity in separating WM and GM for neonates in comparison to conventional relaxation-based contrasts, along with the benefits of fusing multiple contrasts with BISMA. DMRI-derived contrasts depend mainly on micrometer level cellular tissue properties that remain sensitive to tissue differences despite poor infant brain myelination^3-5. Different DTI-derived maps offer complementary information about the underlying tissue types, and such contrasts perform better in segmenting GM and WM in neonate brain compared with non-diffusion weighted image. Segmentation based on single contrast tends to overestimate or under-estimate the GM or WM segmentation, whereas BISMA resolves the over- or under-estimation with segmentations close to ground-truth. Moreover, fusing AD and RD contrasts with BIMSA improves the Dice for segmenting both GM and WM for more than 5%, larger than common inter-rater segmentation error for neonate brain MRI⁷. BISMA will be extended to full infancy age range of 0-2 years old and incorporating diffusion kurtosis imaging (DKI)⁹ or neurite orientation dispersion and density imaging (NODDI)¹⁰contrasts that may further improve the segmentation accuracy.

Acknowledgements

This study is funded by NIH MH092535, MH092535-S1 and HD086984.

References

1. Li G, Wang L, Yap PT, Wang F, Wu Z, Meng Y., …, Shen D. Computational neuroanatomy of baby brains: A review. NeuroImage, 2019;185: 906-925.

2. Zöllei L, Iglesias JE, Ou Y, Grant PE, Fischl B. Infant FreeSurfer: An automated segmentation and surface extraction pipeline for T1-weighted neuroimaging data of infants 0–2 years. Neuroimage, 2020; 218: 116946.

3. Huang H, Shu N, Mishra V, Jeon T, Chalak L, Wang ZJ, ..., He Y. Development of human brain structural networks through infancy and childhood. Cerebral Cortex. 2015; 25.5:1389-1404.

4. Huang H, Zhang J, Wakana S, Zhang W, Ren T, Richards LJ, ..., Mori S. White and gray matter development in human fetal, newborn and pediatric brains. Neuroimage. 2006; 33.1: 27-38.

5. Ouyang M, Dubois J, Yu Q, Mukherjee P, Huang H. Delineation of early brain development from fetuses to infants with diffusion MRI and beyond. Neuroimage. 2019;185: 836-850

6. Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. MICCAI. 2015; pp. 234-241. Springer, Cham.

7. Woo S, Park J, Lee JY, Kweon IS. Cbam: Convolutional block attention module. Proceedings of the European conference on computer vision (ECCV). 2018; 3-19.

8. Weisenfeld NI, Warfield SK. Automatic segmentation of newborn brain MRI. Neuroimage, 2009; 47.2: 564-572.

9. Jensen JH, Helpern JA, Ramani A, Lu H, Kaczynski K. Diffusional kurtosis imaging: the quantification of non‐gaussian water diffusion by means of magnetic resonance imaging. Magn Reson Med: 2005; 53.6: 1432-1440.

10. Zhang H, Schneider T, Wheeler-Kingshott CA, Alexander DC. NODDI: practical in vivo neurite orientation dispersion and density imaging of the human brain. Neuroimage. 2012; 61.4: 1000-1016.

Figures

Figure 1. Poor contrast of T1-weighted (T1w) and T2-weighted (T2w) images on infant brains from 3-6 months and segmentation results from Infant Freesurfer². A-B: Poor T1w and T2w contrasts of a representative 3-6-month infant. C: Infant Freesurfer² segmentation results based on T1w. D: Poor T1w contrast, with yellow contour showing ground-truth cortical boundaries. E: Better T2w contrast in comparison to T1w, with yellow contour showing ground-truth cortical boundaries. F: Over-estimation of cortical boundaries (yellow contour) based on T1w alone by Infant Freesurfer².

Figure 2. Diffusion MRI-derived maps including diffusion tensor imaging (DTI) metrics of a representative neonate subject. A: Ground-truth segmentation along with b0 and fractional anisotropy (FA) from DTI. B: Other DTI contrasts: axial diffusivity (AD), radial diffusivity (RD), and mean diffusivity (MD). Green arrows point out RD and MD have better cortical GM contrasts. Blue arrows point out AD has better deep GM contrast. Yellow arrows indicate the good contrast for myelinated WM on FA.

Figure 3. Example BISMA architecture. Down-sampling path (left) is the encoder; Up-sampling path on is the decoder (right). Convolutional block attention module (CBAM)⁷ is added to each level of the encoder. DoubleConv refers to convolution and rectified linear unit (ReLu) applied successively. Encoder path uses Max-pooling (MaxPool) to down-sample the feature maps, and decoder path uses up-convolution (Up-conv) to up-sample. CBAM-weighted feature maps are concatenated to the decoder. A 1×1×N convolution converts the attention-weighted maps to the output segmentation.

Figure 4. UNet gray matter (GM) and white matter (WM) segmentation results with b0 and DTI contrasts axial, radial, and mean diffusivity (AD, RD, MD) along with BISMA segmentation results. A-B: Individual contrast UNet segmentation results for GM and WM respectively. C-E: Enlarged cortical GM, deep GM, WM areas show the superior performance of BISMA in segmenting different tissue types along with the complementary regional segmentations based on single contrasts F-G: Dice ratio quantification confirms that BIMSA improves both GM (F, Dice=0.93) and WM (G, Dice=0.89) segmentations.

Proc. Intl. Soc. Mag. Reson. Med. 30 (2022)

2159

DOI: https://doi.org/10.58530/2022/2159