Supervised tractogram filtering using Geometric Deep Learning.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Artificial Intelligence

Background:

• Tractograms are virtual representations of brain white matter pathways, crucial for neuroscience research.
• A significant challenge is the presence of non-anatomical fibers (artifacts) in tractograms.
• Current methods for artifact removal often rely on signal reconstruction or topology regularization.

Purpose of the Study:

• To develop and validate Verifyber, a novel method for filtering anatomically implausible fibers from tractograms.
• To leverage anatomical knowledge within a supervised learning framework for improved tractogram accuracy.
• To provide a fast and robust solution for enhancing the quality of white matter representations.

Main Methods:

• A fully-supervised learning approach using a Geometric Deep Learning model named Verifyber.
• Training the model on tractograms annotated according to anatomical principles.
• Employing a sequence Edge Convolution to process each fiber as a graph of points, capturing anatomical properties.
• The model is invariant to fiber orientation and handles variable-sized fibers.

Main Results:

• Verifyber effectively classifies fibers as anatomically plausible or non-plausible.
• The filtering results demonstrate high accuracy and robustness across extensive experiments.
• The method is computationally efficient, filtering 1 million fibers in under a minute on a 12GB GPU.

Conclusions:

• Verifyber offers a significant advancement in tractogram filtering by integrating anatomical knowledge.
• The model provides an accurate, robust, and fast solution for removing non-anatomical fibers.
• This enhances the reliability of tractograms for applications in presurgical planning and understanding brain disorders.