TY - GEN
T1 - Most edges in Markov random fields for white matter hyperintensity segmentation are worthless
AU - Schwarz, Christopher G.
AU - Fletcher, Evan
AU - Singh, Baljeet
AU - Liu, Amy
AU - Smith, Noel
AU - Decarli, Charles
AU - Carmichael, Owen
PY - 2012
Y1 - 2012
N2 - The time and space complexities of Markov random field (MRF) algorithms for image segmentation increase with the number of edges that represent statistical dependencies between adjacent pixels. This has made MRFs too computationally complex for cutting-edge applications such as joint segmentation of longitudinal sequences of many high-resolution magnetic resonance images (MRIs). Here, we show that simply removing edges from full MRFs can reduce the computational complexity of MRF parameter estimation and inference with no notable decrease in segmentation performance. In particular, we show that for segmentation of white matter hyperintensities in 88 brain MRI scans of elderly individuals, as many as 66% of MRF edges can be removed without substantially degrading segmentation accuracy. We then show that removing edges from MRFs makes MRF parameter estimation and inference computationally tractable enough to enable modeling statistical dependencies within and across a larger number of brain MRI scans in a longitudinal series; this improves segmentation performance compared to separate segmentations of each individual scan in the series.
AB - The time and space complexities of Markov random field (MRF) algorithms for image segmentation increase with the number of edges that represent statistical dependencies between adjacent pixels. This has made MRFs too computationally complex for cutting-edge applications such as joint segmentation of longitudinal sequences of many high-resolution magnetic resonance images (MRIs). Here, we show that simply removing edges from full MRFs can reduce the computational complexity of MRF parameter estimation and inference with no notable decrease in segmentation performance. In particular, we show that for segmentation of white matter hyperintensities in 88 brain MRI scans of elderly individuals, as many as 66% of MRF edges can be removed without substantially degrading segmentation accuracy. We then show that removing edges from MRFs makes MRF parameter estimation and inference computationally tractable enough to enable modeling statistical dependencies within and across a larger number of brain MRI scans in a longitudinal series; this improves segmentation performance compared to separate segmentations of each individual scan in the series.
UR - http://www.scopus.com/inward/record.url?scp=84883007920&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84883007920&partnerID=8YFLogxK
U2 - 10.1109/EMBC.2012.6346517
DO - 10.1109/EMBC.2012.6346517
M3 - Conference contribution
C2 - 23366478
AN - SCOPUS:84883007920
SN - 9781424441198
T3 - Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS
SP - 2684
EP - 2687
BT - 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBC 2012
T2 - 34th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS 2012
Y2 - 28 August 2012 through 1 September 2012
ER -