Section: New Results

Biomedical Image Analysis

Participant : Nikos Paragios.

• Image Reconstruction [Nikos Paragios & Hellene Langet]

  In [21] a novel iterative reconstruction algorithm based on compressed sensing was proposed for Digital Subtraction Rotational Angiography (DSRA) that exploits both spatial and temporal sparsity through a proximal implementation that accommodates multiple$L-1$-penalties. These ideas was further explored in [20] where we introduced a three-dimensional reconstruction of tomographic acquisitions in C-arm-based rotational angiography was proposed that was able to deal with the temporal variations due to intra-arterial injections through a compressed-sensing approach leading to significant motion artifacts reduction in spite of the cone-beam geometry, the short-scan acquisition, and the truncated and subsampled data.

• Image Segmentation [Nikos Paragios, Pierre-Yves Baudin, Xiang Bo & Sarah Parisot]

  In [11] the problem of human skeletal muscle segmentation was considered through a graph-based approach (random walker). An automatic seed placement framework was introduced through a graph-theoretic formulation. Towards accounting for anatomical constraints, the Random Walker algorithm was endowed with a liner sub-space statistical prior towards improving segmentation robustness on missing and incomplete data [12] . The same formulation was extended to cope with non-linear priors through a Gaussian-like local prior model penalizing the deviations of the coefficients of the random walker diffusion matrix from the ones learned from the training data [13] . In [25] a novel graph-based prior was considered towards modeling the distribution of low-glioma brain tumors and spatially characterizing them through a sparse hierarchical graph. Such a prior model was integrated to an image-driven voxel-like segmentation framework where image separation was achieved through a machine learning method towards automatic detection, characterization and segmentation of brain tumors. Furthermore, towards encoding pose invariance in the context of knowledge-based segmentation in [33] where a higher order graph-based implicit pose invariant formulation was introduced for cardiac segmentation. The formulation was endowed with higher order cliques allowing (i) the estimation of boundary and regional image support and (ii) the implicit modeling of local deformations with respect to a prior statistical model while being invariant to linear transformations.

• Image Registration [Nikos Paragios, Nicolas Honnorat & Sarah Parisot]

  In [15] the problem of organ-driven registration was addressed through simultaneous combined fusion of multi-modal images in the context of guide-wire segmentation through fluoroscopic and contrast enhanced images. To this end, a graphical model was considered that was segmenting and registering the guide-wire in the two modalities while establishing correspondences between the associated curves as well. Similar philosophy was used in the [26] where a method for one shot deformable brain registration and tumor segmentation was proposed between a healthy anatomical atlas and a diseased patient. Both tasks were addressed through a discrete formulation (pair-wise MRF using grid-like deformation models and machine learning discriminative frameworks for the separation of healthy versus diseased tissues) while interconnections between the two graphs were used to alleviate the registration requirement on tumor areas. The problem of symmetric registration was studied in [31] through a common grid deforming in both directions according to a symmetric manner towards minimizing the image similarity criterion between the source and the target image while guaranteeing the expected diffeomorphic nature of the deformation field.

• Computational Anatomy [Nikos Paragios]

  In [16] we introduced a novel approach for detecting the presence of white matter lesions in periventricular areas of the brain using manifold-constrained embeddings. The proposed method uses locally linear embedding (LLE) to create ”normality” distributions of the brain where deviations from the manifolds are estimated by calculating geodesic distances along locally linear planes in the embedding. Experiments highlight the need of nonlinear techniques to learn the studied data leading to outstanding detection rates when comparing individuals to a specific pathological pattern.