Entrack: Probabilistic Spherical Regression with Entropy Regularization for Fiber Tractography

• Published in: International journal of computer vision Vol. 129; no. 3; pp. 656 - 680
• Main Authors: Wegmayr, Viktor, Buhmann, Joachim M.
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2021; Springer; Springer Nature B.V
• Subjects: Artificial Intelligence; Computer Imaging; Computer Science; Image Processing and Computer Vision; In vivo methods and tests; Information theory; Machine learning; Magnetic resonance imaging; Maximum entropy; Noise levels; Pattern Recognition; Pattern Recognition and Graphics; Probabilistic methods; Probabilistic models; Regression models; Regularization; Special Issue: Computer Vision and Pattern Recognition; Statistical analysis; Vision; Brain; Diffusion MRI; Algorithm validation; Maximum-entropy inference; Machine Learning; Tractography
• ISSN: 0920-5691; 1573-1405; 1573-1405
• DOI: 10.1007/s11263-020-01384-1

White matter tractography, based on diffusion-weighted magnetic resonance images, is currently the only available in vivo method to gather information on the structural brain connectivity. The low resolution of diffusion MRI data suggests to employ probabilistic methods for streamline reconstruction, i.e., for fiber crossings. We propose a general probabilistic model for spherical regression based on the Fisher-von-Mises distribution, which efficiently estimates maximum entropy posteriors of local streamline directions with machine learning methods. The optimal precision of posteriors for streamlines is determined by an information-theoretic technique, the expected log-posterior agreement concept. It relies on the requirement that the posterior distributions of streamlines, inferred on retest measurements of the same subject, should yield stable results within the precision determined by the noise level of the data source.