Learning Streamline Embeddings with Variational Autoencoder for Intersubject Bundle Comparison in Alzheimer’s Disease

Background Tractograms generated from diffusion MRI (dMRI) can be used to evaluate white matter abnormalities in Alzheimer's Disease patients (AD). However, the large quantity of streamlines causes problems for downstream analysis. Here we use a variational autoencoder (VAE) with a 1D convolutional layer (ConvVAE) to embed streamlines in a 2D latent space. The generative nature of ConvVAE allows us to apply Euclidean distances directly to embeddings and perform intersubject fiber bundle comparisons. Method Multi‐shell dMRI from 141 subjects ‐ 87 cognitively normal controls (CN), 44 with mild cognitive impairment (MCI), 10 with dementia (AD) ‐ from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) were preprocessed using the ADNI3 dMRI protocol. 30 white matter tracts were extracted from whole brain tractograms per subject using DIPY's RecoBundles. A control subject with 51,654 streamlines was used to train the ConvVAE with Evidence lower bound (ELBO) loss. We evaluated results by visualizing 2D embeddings and reconstructed streamlines. K‐nearest neighbor (KNN) clustering with Euclidean distance (k=5) was performed on the control subject in a bundle labeling task, to evaluate the alignment between embeddings and generated bundle labels. Result 2D embeddings for the control subject used for training and a randomly selected MCI and AD subject are shown in Fig. 1, colored by bundles (top) and hemisphere (bottom). Embeddings for each bundle are aligned between subjects, reflecting the shape, size and orientation of the bundle in the streamline space. Embeddings, streamlines and reconstruction for 4 left hemisphere bundles are plotted in Fig. 2. Linearly interpolated points sampled from the latent space translate into smooth transitions in the streamline space, so Euclidean distance can be used in downstream tasks. The fitted KNN model was evaluated on all other subjects where the weighted accuracy for each group was 80.56% (CN), 78.59% (MCI) and 75.76% (AD). Conclusion ConvVAE generates 2D embeddings that preserve bundles' spatial and shape information. It learns a smooth latent space from streamlines, which allows for meaningful decodings from sampled points and can be directly applied to new data. Distance‐based algorithms can be used in downstream tasks, such as bundle labeling and intersubject comparisons.