Deep Learning-Based Segmentation of Brainstem White Matter Using Tractographic Guidance

1606 

Abstract Submission 

Mark Olchanyi^1,2, Schreier David^1,2,3,4, Hannah Kinney^5,3, Juan Eugenio Iglesias^6,2,3, Emery Brown^1,2,3, Brian Edlow^6,2,3

¹Massachusetts Institute of Technology, Cambridge, MA, ²Massachusetts General Hospital, Boston, MA, ³Harvard Medical School, Boston, MA, ⁴University of Bern, Bern, Switzerland, ⁵Boston Children's Hospital, Boston, MA, ⁶Athinoula A. Martinos Center for Biomedical Imaging, Charlestown, MA

Mark Olchanyi  
Massachusetts Institute of Technology|Massachusetts General Hospital
Cambridge, MA|Boston, MA

Schreier David, MD, PhD  
Massachusetts Institute of Technology|Massachusetts General Hospital|Harvard Medical School|University of Bern
Cambridge, MA|Boston, MA|Boston, MA|Bern, Switzerland

Hannah Kinney  
Boston Children's Hospital|Harvard Medical School
Boston, MA|Boston, MA

Juan Eugenio Iglesias, Ph.D.  
Athinoula A. Martinos Center for Biomedical Imaging|Massachusetts General Hospital|Harvard Medical School
Charlestown, MA|Boston, MA|Boston, MA

Emery Brown  
Massachusetts Institute of Technology|Massachusetts General Hospital|Harvard Medical School
Cambridge, MA|Boston, MA|Boston, MA

Brian Edlow, M.D.  
Athinoula A. Martinos Center for Biomedical Imaging|Massachusetts General Hospital|Harvard Medical School
Charlestown, MA|Boston, MA|Boston, MA

Diffusion MRI (dMRI) provides a non-invasive way of studying white matter that connects regions of interest (ROIs) in the brain (Basser, 2000) (Mori, 2001). Reliably reconstructing brainstem white matter bundles (WMB) in dMRI is crucial to the study and treatment of diseases of the central nervous system (CNS) with subcortical pathology. However, brainstem WMB segmentation methods are lacking due to the brainstem's small size and architectural complexity. We present an automated brainstem WMB segmentation method, the BrainStem Bundle Tool (BSB). BSB segmentation is performed on a tract intensity map (SIM) that is formed by passing WM streamlines between regions that are structurally adjacent to the rostral brainstem and are readily segmentable. For the segmentation model, we utilize a Convolutional Neural Network (CNN) architecture that is modified with two distinct elements that emphasize small-structure segmentation: an attention gate and a semi-dense conditional random field (CRF) attached to the CNN output, which inflates posterior probabilities around high WMB probability regions.

We identified seven candidate brainstem WMB that are visualizable across both ex vivo and clinical-resolution in vivo dMRI volumes. A U-Net CNN (Ronneberger, 2015) with a fine attention gate (Schlemper, 2019) positioned between the highest-resolution encoder and decoder layers is used, which inputs low-b, fractional anisotropy, and SIM volumes. Each SIM volume is created via probabilistically propagating streamlines between four ROIs adjacent to the ponto-mesencephalic component of the brainstem: the medulla, cerebellar grey matter, thalamus, and ventral diencephalon (Figure 1 A/B). All inputs are resampled to 1mm isotropic resolution and centered around the center-of-mass of a pontine mask. The SoftMax CNN layer dilutes posteriors of thin WMB in low-resolution and ex vivo dMRI modalities, therefore we employ label posterior refinement via a semi-dense CRF (Figure 1C). The CNN was trained on 30 subjects from the WU-Minn Human Connectome Project dMRI dataset (Van Essen, 2013). The training set is aggressively augmented to mimic multiple dMRI modalities, including shell dropout randomization, resampling, noise injection, and local deformation.
We tested BSB segmentation accuracy on 10 HCP subjects at 1.25mm isotropic resolution, 7 ex vivo specimens scanned at 0.75mm isotropic resolution, and 15 control subjects from the Alzheimer's Disease Neuroimaging Initiative 3 (ADNI3) dataset at 2mm isotropic resolution (Wiener, 2017) (Figure 2A). We then proceeded to assess the translatability of BSB via an along-tract diffusivity metric classification task between ADNI3 Alzheimer's (AD) patients and control subjects (Figure 2B/C).

BSB displayed robust WMB segmentation accuracy across all three datasets (HCP, ex vivo, and ADNI3) (Figure 1A). Furthermore, inclusion of the attention gate, CRF, and the SIM channel yielded significantly higher accuracies across both the ADNI3 control and ex vivo datasets. In AD classification, BSB WMB showed a similar discriminatory power to cortical WMB, and superior discriminatory power to a cortical white matter mask and the whole brainstem in when assessing average along-tract mean diffusivity classification between cognitively normal controls and AD/mildly cognitively impaired subjects in the ADNI3 dataset (Figure 2B/C).

We present BSB, a tool that creates a tractographic mapping of streamline intensities corresponding to brainstem WMB from extra-brainstem regions and provides candidate WMB segmentations from those mappings. BSB accurately segments brainstem WMB in both high-resolution ex vivo and low-resolution in vivo dMRI data. We show that one translational application of BSB is in Alzheimer's disease classification via assessment of brainstem WMB diffusivity metrics. We anticipate that subcortical WMB segmentation with BSB can be used to elucidate subcortical white matter pathology in a broad spectrum of CNS diseases.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)

Connectivity (eg. functional, effective, structural)

Diffusion MRI Modeling and Analysis ¹

Segmentation and Parcellation ²

Anatomical MRI

Brainstem

Degenerative Disease

Machine Learning

MRI

Structures

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC