Section: New Results

Biomedical Image Analysis

In [38] an overview of the methodological foundations of biomedical image analysis as well as their use to provide answers to a variety of clinical problems are presented. The problem of volumes of rotational angiography using non-linear sparsity constraints was studied in [28] where a novel method able to handle highly under-sampled acquisitions was introduced.

[18] presents an overview of the use of graphical models in artificial vision where both inference, learning as well as applications are discussed. In [32] a max-margin dual decomposition method was used towards learning the compact, pose invariant shape representation using higher order graphs acting both on the connectivity of the graph as well its potentials. Graphical model was used as prior in [13] under a "curve" propagation principle for generic prior-constrained organ segmentation in 2D images. Similar inspiration driven from a higher order pose invariant graphical model learned according to [32] was considered in [31] where a novel segmentation method was proposed coupling model-based and pixel-based concepts while being pose invariant. The underlying idea was to consider a two-layer interconnected graphical model acting on pixel and on control points where segmentation consistency was imposed throgh penalties on label discrepancies of the different layers. Higher order graphical models were also employed in [14] for spine segmentation using an articulated graphical model where a non-linear approach/embedding towards reducing the complexity of the inference step was considered at training.

In [16] a comprehensive survey of deformable registration was presented. It was organized in three sections: the first was studying the deformation model, the second the similarity criterion while the last section discussed the different optimization strategies. The problem of atlas-based segmentation/registration in the presence of brain tumors was studied in [29] an adaptive uncertainty-driven sampling strategy was proposed coupling segmentation and registration. Both sampling spaces (quantization of the search space, deformation grid) were determined according to the observed optimization min-marginals. The challenging problem of image to slice registration was proposed in [23] where an over-parameterized low rank graphical model acting both on the plan selection as well the in-plane deformations was introduced. The main strength of the method was its ability to simultaneously recover both the plane and the organ deformation.