A Joint Bayesian Framework for MR Brain Scan Tissue and Structure Segmentation Based on Distributed Markovian Agents - Computational Intelligence in Healthcare: Advanced Methodologies - page 93

Fig. 7. Evaluation of FBM-TS on IBSR v2 (9 right structures) and comparison with
[27]. Y axis = Mean Dice metric
84%, 91% for caudate, putamen and thalamus. We then considered 18 images
from the IBSR v2 database. The mean Dice metric for the 9 right structures (17
were segmented) is reported in Figure 7.
7 Discussion
The results obtained with our approach are very satisfying and compare favor-
ably with other existing methods. The strength of our fully Bayesian joint model
is to be based on the specification of a coherently linked system of conditional
models for which we make full use of modern statistics to ensure tractability. The
tissue and structure models are linked conditional MRF's that capture several
level of interactions. They incorporate 1) spatial dependencies between voxels for
robustness to noise, 2) relationships between tissue and structure labels for co-
operative aspects and 3) apriori anatomical information via the MRF external
field parameters for consistency with expert knowledge.
Besides, the addition of a conditional MRF model on the intensity distribution
parameters allows us to handle local estimations for robustness to nonuniformi-
ties. In this setting, the whole consistent treatment of MR brain scans is made
possible using the framework of Generalized Alternating Minimization (GAM)
procedures that generalize the standard EM framework. Another advantage of
this approach is that it is made of steps that are easy to interpret and could be
enriched with additional information. In particular, results currently highly de-
pend on the atlas registration step which could be introduced in our framework
as in [28]. A step in this direction is proposed in [29]. Also a different kind of prior
knowledge could be considered such as the fuzzy spatial relations used in [4].
Other on going work relates to the interpolation step we added to increase
robustness to nonuniformities at a voxel level. We believe this stage could be
generalized and incorporated in the model by considering successively various