Meeting Banner
Abstract #0782

Automatic Multi-Label Segmentation of the Preterm Brain with the Use of Adaptive Atlases

Antonios Makropoulos1, Ioannis S. Gousias1, Christian Ledig2, Paul Aljabar2, Ahmed Serag2, Joseph V. Hajnal3, A. David Edwards1, Serena J. Counsell1, Daniel Rueckert2

1Centre for the Developing Brain, Imperial College London, London, United Kingdom; 2Biomedical Image Analysis Group, Department of Computing, Imperial College London, London, United Kingdom; 3Imaging Sciences Department, MRC Clinical Sciences Centre, Imperial College London, London, United Kingdom

The changes that occur in brain anatomy between early preterm and term age are significant and present challenges to an accurate automatic MRI segmentation of the preterm brain. Atlas-based techniques are amongst the most popular in brain MRI segmentation. However, when the atlases used deviate significantly from the subjects to be segmented, they result in inaccuracies. Within this context, we propose a Expectation-Maximization segmentation technique to propagate multiple labels, 50 in total, from manually segmented atlases around term to early preterm ages. The presented framework further adapts the priors to the target image so as to compensate for these inaccuracies.

Abstract PDF Poster PDF

Keywords

according account accurate achieved adaptation adapted adaptive allow allows alternatively anatomy around arrows assessment assisted atlas atlases authoritatively automatic available background blinded borders brain brains challenge clearly clinical college combined compensate computer computing constraints corrected correction cortex cortical defined describe described developing development differentiation divide divided dividing early employing encompasses equivalent errors expert field fluid form framework fusion graph gray grossly growth imperial important imposes improvement inaccuracy included incorrectly increasingly infants initial intensities intensity intervention kingdom label labels late latter limitation manual manually many median medical mixture model models neonatal neonates onto overlaid partial particularly population presence prior priors probabilistic problematic process processing propagated proposed protocol purple random ranging rapid refined regional registered registration rely removed removes scheme sciences segmentation segmentations segmented segmenting sets shape significantly simple song source space spatial structure structures subject subjects target term thesis tissue transactions transformed types uniformity ventricles visible volume vulnerable weeks white yield younger