Docteur de l'université Automatic Segmentation and Shape Analysis ...

28 Chapter 2 Literature Review
determined by consensus of the voting (Heckemann et al., 2006). A more sophis-
ticated approach is to weight the votes from each atlas based on their perfor-
mance, such as in simultaneous truth and performance level estimation (STAPLE,
Warfield et al., 2004) which estimates the performance level of the atlases using an
expectation-maximization (EM) algorithm. In contrast to STAPLE, in which the
weights are globally evaluated on the atlas, locally weighted fusion (Commowick
et al., 2009; van Rikxoort et al., 2010; Coupé et al., 2010) were developed to reduce
the error in the lower ranked atlases by selecting the regions or voxels locally more
similar to the target image. Utilizing the intensity information of the target and
the atlas images in the fusion step, the local weighted voting (Artaechevarria et al.,
2009) method weights each atlas in the voting based on their local similarity to the
target. Both global and local weighting are special cases under a generative proba-
bilistic framework developed by Sabuncu et al. (2010). In general, local weighting
achieves higher accuracy than global weighting. Applying MRF smoothing to the
STAPLE results may improve the performance of the fusion (Leung et al., 2010).
Statistical learning based methods are introduced in the multi-atlas based seg-
mentation to infer the performance of each atlas and to weight the atlases. The
segmentation accuracy map (Sdika, 2010) estimates the average accuracy of each
voxel in the atlas on the training set, and this estimation is used as the weight
in the label fusion. In the fusion by supervised learning and dynamic information
(SuperDyn, Khan et al., 2011), the accuracy weights are learned by a linear regres-
sion of segmentation accuracy with Tikhonov regularization, which is combined
with local estimates of registration accuracy most similar to that of Artaechevarria
et al. (2009).
The multi-atlas segmentation has also been developed in combination with ap-
proaches. In the work of van der Lijn et al. (2008), the multi-atlas method is used
to generate the probability map as the spatial prior, which is then combined with
the intensity model and neighborhood model to formulate the energy function.
The structures are segmented by minimizing the energy function using a graph
cut algorithm (Kolmogorov and Zabin, 2004). The intensity model is refined by