Multi-atlas based segmentation of brain images: atlas selection and its effect on accuracy.

Visual Information Processing Group, Department of Computing, Imperial College London, 180 Queen's Gate, London, SW7 2AZ, UK.
NeuroImage (Impact Factor: 6.13). 03/2009; 46(3):726-38. DOI: 10.1016/j.neuroimage.2009.02.018
Source: PubMed

ABSTRACT Quantitative research in neuroimaging often relies on anatomical segmentation of human brain MR images. Recent multi-atlas based approaches provide highly accurate structural segmentations of the brain by propagating manual delineations from multiple atlases in a database to a query subject and combining them. The atlas databases which can be used for these purposes are growing steadily. We present a framework to address the consequent problems of scale in multi-atlas segmentation. We show that selecting a custom subset of atlases for each query subject provides more accurate subcortical segmentations than those given by non-selective combination of random atlas subsets. Using a database of 275 atlases, we tested an image-based similarity criterion as well as a demographic criterion (age) in a leave-one-out cross-validation study. Using a custom ranking of the database for each subject, we combined a varying number n of atlases from the top of the ranked list. The resulting segmentations were compared with manual reference segmentations using Dice overlap. Image-based selection provided better segmentations than random subsets (mean Dice overlap 0.854 vs. 0.811 for the estimated optimal subset size, n=20). Age-based selection resulted in a similar marked improvement. We conclude that selecting atlases from large databases for atlas-based brain image segmentation improves the accuracy of the segmentations achieved. We show that image similarity is a suitable selection criterion and give results based on selecting atlases by age that demonstrate the value of meta-information for selection.

  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: A new TRIO algorithm method integrating three different algorithms is proposed to perform brain MRI segmentation in the native coordinate space, with no need of transformation to a standard coordinate space or the probability maps for segmentation. The method is a simple voxel-based algorithm, derived from multispectral remote sensing techniques, and only requires minimal operator input to depict GM, WM, and CSF tissue clusters to complete classification of a 3D high-resolution multislice-multispectral MRI data. Results showed very high accuracy and reproducibility in classification of GM, WM, and CSF in multislice-multispectral synthetic MRI data. The similarity indexes, expressing overlap between classification results and the ground truth, were 0.951, 0.962, and 0.956 for GM, WM, and CSF classifications in the image data with 3% noise level and 0% non-uniformity intensity. The method particularly allows for classification of CSF with 0.994, 0.961 and 0.996 of accuracy, sensitivity and specificity in images data with 3% noise level and 0% non-uniformity intensity, which had seldom performed well in previous studies. As for clinical MRI data, the quantitative data of brain tissue volumes aligned closely with the brain morphometrics in three different study groups of young adults, elderly volunteers, and dementia patients. The results also showed very low rates of the intra- and extra-operator variability in measurements of the absolute volumes and volume fractions of cerebral GM, WM, and CSF in three different study groups. The mean coefficients of variation of GM, WM, and CSF volume measurements were in the range of 0.03% to 0.30% of intra-operator measurements and 0.06% to 0.45% of inter-operator measurements. In conclusion, the TRIO algorithm exhibits a remarkable ability in robust classification of multislice-multispectral brain MR images, which would be potentially applicable for clinical brain volumetric analysis and explicitly promising in cross-sectional and longitudinal studies of different subject groups.
    PLoS ONE 01/2015; 10(2):e0115527. DOI:10.1371/journal.pone.0115527 · 3.53 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Multi-atlas based methods have been a trend for robust and automated image segmentation. In general these methods first transfer prior manual segmentations, i.e., label maps, on a set of atlases to a given target image through image registration. These multiple label maps are then fused together to produce segmentations of the target image through voting strategy or statistical fusing, e.g., STAPLE. STAPLE simultaneously estimates the true segmentation and the label map performance level, but has been shown inaccurate for multi-atlas segmentation because it is determined completely on the propagated label maps without considering the target image intensity. We develop a new method, called iSTAPLE, that combines target image intensity into a similar maximum likelihood estimate (MLE) framework as in STAPLE to take advantage of both intensity-based segmentation and statistical label fusion based on atlas consensus and performance level. The MLE framework is then solved using a modified EM algorithm to simultaneously estimate the intensity profiles of structures of interest as well as the true segmentation and atlas performance level. Unlike other methods, iSTAPLE does not require the target image to have same image contrast and intensity range as the atlas images, which greatly extends the use of atlases. Experiments on whole brain segmentation showed that iSTAPLE performed consistently better than STAPLE.
    SPIE Medical Imaging; 03/2013
  • [Show abstract] [Hide abstract]
    ABSTRACT: Multi-atlas segmentation methods are among the most accurate approaches for the automatic labeling of magnetic resonance (MR) brain images. The individual segmentations obtained through multi-atlas propagation can be combined using an unweighted or locally weighted fusion strategy. Label overlaps can be further improved by refining the label sets based on the image intensities using the Expectation-Maximisation (EM) algorithm. A drawback of these approaches is that they do not consider knowledge about the statistical intensity characteristics of a certain anatomical structure, especially its intensity variance. In this work we employ learned characteristics of the intensity distribution in various brain regions to improve on multi-atlas segmentations. Based on the intensity profile within labels in a training set, we estimate a normalized variance error for each structure. The boundaries of a segmented region are then adjusted until its intensity characteristics are corrected for this variance error observed in the training sample. Specifically, we start with a high-probability “core” segmentation of a structure, and maximise the similarity with the expected intensity variance by enlarging it. We applied the method to 35 datasets of the OASIS database for which manual segmentations into 138 regions are available. We assess the resulting segmentations by comparison with this gold-standard, using overlap metrics. Intensity-based statistical correction improved similarity indices (SI) compared with EM-refined multi-atlas propagation from 75.6% to 76.2% on average. We apply our novel correction approach to segmentations obtained through either a locally weighted fusion strategy or an EM-based method and show significantly increased similarity indices.
    SPIE Medical Imaging; 03/2013

Full-text (2 Sources)

Available from
Jun 3, 2014