DeepMultiConnectome: Deep multi-task prediction of structural connectomes directly from diffusion MRI tractography.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Diffusion MRI (dMRI) tractography maps in vivo brain structural connections.
• Traditional connectome generation is slow and requires gray matter parcellation, limiting large-scale studies.

Purpose of the Study:

• Introduce DeepMultiConnectome, a deep-learning model for direct structural connectome prediction from tractography.
• Enable efficient and flexible generation of subject-specific connectomes supporting multiple parcellation schemes.

Main Methods:

• Utilize a point-cloud-based neural network with multi-task learning to classify streamlines.
• Train and validate the model on Human Connectome Project data with 84 and 164 region parcellations.
• Predict connectomes from tractograms in approximately 40 seconds.

Main Results:

• DeepMultiConnectome predicts structural connectomes with high agreement to traditional methods across two parcellation schemes.
• Achieved high Pearson correlations (r=0.727-0.995) for various weighting strategies (streamline-count, SIFT2, mean-FA).
• Demonstrated comparable performance and reproducibility to traditional methods in test-retest and downstream prediction tasks.

Conclusions:

• DeepMultiConnectome offers a fast, scalable solution for generating subject-specific connectomes.
• The model supports multiple parcellation and weighting schemes, enhancing flexibility for connectome research.
• Facilitates large-scale dMRI studies by overcoming limitations of traditional connectome generation methods.