Details

Autor(en) / Beteiligte

• Alam, Saruar
• Halandur, Bharath
• Mana, P. G. L. Porta
• Goplen, Dorota
• Lundervold, Arvid
• Lundervold, Alexander Selvikvåg

Titel

Brain Tumor Segmentation from Multiparametric MRI Using a Multi-encoder U-Net Architecture

Ist Teil von

• Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, p.289-301

Ort / Verlag

Cham: Springer International Publishing

Link zum Volltext

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• This paper describes our submission to Task 1 of the RSNA-ASNR-MICCAI Brain Tumor Segmentation (BraTS) Challenge 2021, where the goal is to segment brain glioblastoma sub-regions in multi-parametric MRI scans. Glioblastoma patients have a very high mortality rate; robust and precise segmentation of the whole tumor, tumor core, and enhancing tumor subregions plays a vital role in patient management. We design a novel multi-encoder, shared decoder U-Net architecture aimed at reducing the effect of signal artefacts that can appear in single channels of the MRI recordings. We train multiple such models on the training images made available from the challenge organizers, collected from 1251 subjects. The ensemble-model achieves Dice Scores of 0.9274±0.0930\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.9274 \pm 0.0930$$\end{document}, 0.8717±0.2456\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.8717 \pm 0.2456$$\end{document}, and 0.8750±0.1798\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.8750 \pm 0.1798$$\end{document}; and Hausdorff distances of 4.77±17.05\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.77 \pm 17.05$$\end{document}, 17.97±71.54\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$17.97 \pm 71.54$$\end{document}, and 10.66±55.52\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10.66 \pm 55.52$$\end{document}; for whole tumor, tumor core, and enhancing tumor, respectively; on the 570 test subjects assessed by the organizer. We investigate the robustness of our automated segmentation system and discuss its possible relevance to existing and future clinical workflows for tumor evaluation and radiation therapy planning.