Filtering in tractography using autoencoders (FINTA)

Abstract

Current brain white matter fiber tracking techniques show a number of problems, including: generating large proportions of streamlines that do not accurately describe the underlying anatomy; extracting streamlines that are not supported by the underlying diffusion signal; and under-representing some fiber populations, among others. In this paper, we describe a novel unsupervised learning method to filter streamlines from diffusion MRI tractography, and hence, to obtain more reliable tractograms.We show that a convolutional neural network autoencoder provides a straightforward and elegant way to learn a robust representation of brain streamlines, which can be used to filter undesired samples with a nearest neighbor algorithm. Our method, dubbed FINTA (Filtering in Tractography using Autoencoders) comes with several key advantages: training does not need labeled data, as it uses raw tractograms, it is fast and easily reproducible, it does not rely on the input diffusion MRI data, and thus, does not suffer from domain adaptation issues. We demonstrate the ability of FINTA to discriminate between "plausible" and "implausible" streamlines as well as to recover individual streamline group instances from a raw tractogram, from both synthetic and real human brain diffusion

Microstructural Mapping of Neural Pathways in Alzheimer's Disease using Macrostructure-Informed Normative Tractometry

“…It is also publicly available, preprocessed, and carefully quality-controlled. However, in the future, we will assess tractography data from multi-shell acquisitions such as the Human Connectome Project (HCP) 86 in the pre-training stage to better learn the organization of complex fiber populations, and apply tractogram filtering techniques such as SIFT/SIFT2 87,88 , COMMIT/COMMIT2 89,90 and FINTA 38 to reduce the proportion of false positive streamlines. Additionally, diffusion MRI measures derived from multi-shell data, such as neurite orientation dispersion and density imaging (NODDI) 64 , diffusion kurtosis imaging (DKI) 91 , and MAP-MRI 92 may provide a richer feature set and may detect AD-related abnormalities with greater sensitivity and specificity 4,93,94 .…”

“…In MINT, we use 1D convolutional layers to model streamline-level information. This method was proposed to enforce sequential dependency of 3D points along a streamline 38 . Illustrated in Figure 2(a), a 1D convolution learns from local neighbor-hood points along a streamline using a ‘sliding window’, and multiple convolutional layers can be stacked to learn information from points farther away.…”

Variational Autoencoders for Generating Synthetic Tractography-Based Bundle Templates in a Low-Data Setting

“…Bundle labels were not used in training, and only the streamlines' coordinates were passed in as model input, where one streamline is one sample. Each streamline was resampled to 256 equidistant points to allow the use of convolutional layers [23]. All streamlines from each subject were normalized to fit into a standard sphere, where the centroid and radius are calculated from the atlas data (https://figshare.com/articles/dataset/ Atlas_of_30_Human_Brain_Bundles_in_MNI_ space/12089652) [12] to account for any misalignment between subjects.…”

“…Prior studies have investigated the use of autoencoders, an architecture that encodes high dimensional data into a low dimensional latent space and optimizes the mapping to minimize data reconstruction error, on tractography data. FINTA [23] uses a streamline-based autoencoder to learn from a labeled tractogram for streamline filtering. However, training the model on one tractogram is limited in capturing large shape variations across subjects, and the processing of labeling each streamline in the raw tractogram can be time-consuming.…”

“…Prior studies have investigated the use of autoencoders, an architecture that encodes high dimensional data into a low dimensional latent space and optimizes the mapping to minimize data reconstruction error, on tractography data. FINTA [23], GESTA [24] ad FIESTA [25] uses streamline-based autoencoders to learn from tractogram(s) for streamline filtering, generative sampling in the latent space and tractogram segmentation. However, training on whole-brain tractograms which contains a large amount of streamlines makes it more difficult to capture variations across subjects.…”