gms | German Medical Science

70. Jahrestagung der Deutschen Gesellschaft für Neurochirurgie (DGNC)
Joint Meeting mit der Skandinavischen Gesellschaft für Neurochirurgie

Deutsche Gesellschaft für Neurochirurgie (DGNC) e. V.

12.05. - 15.05.2019, Würzburg

Automated tract segmentation in language tumour patients

Automatisierte Faserbahnsegmentierung bei Sprachtumorpatienten

Meeting Abstract

Search Medline for

  • presenting/speaker Lucius Fekonja - Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
  • Ziqian Wang - Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
  • Peter Vajkoczy - Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
  • Thomas Picht - Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland

Deutsche Gesellschaft für Neurochirurgie. 70. Jahrestagung der Deutschen Gesellschaft für Neurochirurgie (DGNC), Joint Meeting mit der Skandinavischen Gesellschaft für Neurochirurgie. Würzburg, 12.-15.05.2019. Düsseldorf: German Medical Science GMS Publishing House; 2019. DocV284

doi: 10.3205/19dgnc303, urn:nbn:de:0183-19dgnc3035

Published: May 8, 2019

© 2019 Fekonja et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 License. See license information at http://creativecommons.org/licenses/by/4.0/.


Outline

Text

Objective: White matter tractography is a challenge in brain tumor patients. Especially tractography of language related fiber bundles is time consuming and difficult. Furthermore, anatomical seeding for tractography is subjective, influenced by the a-priori knowledge and expectations of the examiner. Novel machine learning approaches enabling automated tract segmentation open up the possibility to overcome this shortcoming. Here, we tested for the first time the feasibility of automated tract segmentation in brain tumor patients.

Methods: We included 47 patients with gliomas grade III and IV in the left hemisphere and invasion of the language network. The open source package TractSeg provides a convolutional neural network-based approach that directly segments tracts without tractography. It uses a semi-automatic approach to segment reference segmentations of 72 anatomically clearly described tracts in a cohort of 105 selected subjects from the Human Connectome Project (HCP). Subsequently, tractography and along-tract statistical analysis can be performed based on prior segmented fiber bundles.

Results: Automated tract segmentation proved to be feasible in 33/47 brain tumor patients, generated on routine clinical data. We segmented 72 distinct fiber bundles per patient with the open source package, resulting in 2376 detailed tractographies. In 14 patients no feasible tracts could be automatically generated. This failure was less due to the impact of pathology on white matter integrity, but more to wrong calculations of complex within-voxel modelling of fiber orientation distributions.

Conclusion: Machine learning based automated tract segmentation is already feasible in brain tumor patients with routine MRI data quality. Yet, the algorithm is currently still prone to errors, necessitating manual correction by the examiner. With further improvement of the algorithmic analyses, automated tract segmentation has the potential to revolutionize tractography in clinical daily life.