Generative Sampling in Bundle Tractography using Autoencoders (GESTA).

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Current tractography methods rely on local orientation, often resulting in incomplete reconstructions and poor spatial coverage of white matter pathways.
• Certain white matter bundles are inherently difficult to track, leading to gaps and inaccuracies in brain connectivity mapping.

Purpose of the Study:

• To introduce GESTA (Generative Sampling in Bundle Tractography using Autoencoders), a novel deep generative method for improving white matter tractography.
• To enhance the spatial coverage and completeness of streamline reconstructions, particularly for hard-to-track bundles.

Main Methods:

• Developed a generative, autoencoder-based framework (GESTA) for bundle-wise streamline generation.
• GESTA does not require propagation of local orientations and uses a single model for generating streamlines.
• Streamline evaluation framework assesses anatomical plausibility, local orientation alignment, and geometry features, including optional gray matter connectivity.

Main Results:

• GESTA significantly improves white matter volume coverage in poorly populated bundles on both synthetic and in vivo human brain data.
• Demonstrated substantial gains in bundle overlap, with 1.5x improvement on "Fiber Cup" and 6x on ISMRM 2015 Tractography Challenge datasets.
• Achieved a 4x white matter volume increase on the BIL&GIN callosal homotopic dataset and successfully populated bundles in the TractoInferno dataset.

Conclusions:

• GESTA is a novel deep generative bundle tractography method that effectively improves white matter reconstruction.
• The autoencoder-based approach offers superior spatial coverage and completeness compared to traditional and other deep learning methods.
• GESTA provides a powerful tool for enhancing the accuracy and detail of brain white matter pathway mapping.