Section: New Results

Image Segmentation, Registration and Analysis

Non Local Means-based Speckle Filtering for Ultrasound Images

Participants : Pierrick Coupé, Pierre Hellier, Christian Barillot.

In ultrasound (US) imaging, preprocessing is expected to improve the performance of quantitative image analysis techniques. In this work [15] , an adaptation of the Non Local (NL) means filter is proposed to denoise ultrasound images. Originally developed for additive white Gaussian noise, we propose a Bayesian framework to design an NL-means filter adapted to a relevant ultrasound noise model. 2D and 3D experiments were carried out on synthetic and real images. Quantitative results on synthetic images with various noise models demonstrate that the proposed method outperforms state-of-the-art methods for speckle reduction. Results on real images show that the proposed method is very efficient in terms of edge preservation and noise removal. Finally, we introduced a new registration-based evaluation framework and we show that the NL-means-based speckle filter is very competitive to accurately register real images, compared to other denoising methods.

Prior affinity measures on matches for ICP-like nonlinear registration of free-form surfaces

Participants : Benoît Combès, Sylvain Prima.

In this work, we showed that several well-known nonlinear surface registration algorithms can be put in an ICP-like framework, and thus boil down to the successive estimation of point-to-point correspondences and of a transformation between the two surfaces. We proposed to enrich the ICP-like criterion with additional constraints and showed that it is possible to minimise it in the same way as the original formulation, with only minor modifications in the update formulas and the same convergence properties. These constraints help the algorithm to converge to a more realistic solution and can be encoded in an affinity term between the points of the surfaces to register. This term is able to encode both a priori knowledge and higher order geometrical information in a unified manner. We illustrated the high added value of this new term on synthetic and real data [35] .

Setting priors and enforcing constraints on matches for nonlinear registration of meshes

Participants : Benoît Combès, Sylvain Prima.

In this work, we showed that a simple probabilistic modelling of the registration problem for surfaces allows to solve it by using standard clustering techniques. In this framework, point-to-point correspondences are hypothesized between the two free-form surfaces, and we showed how to specify priors and to enforce global constraints on these matches with only minor changes in the optimisation algorithm. The purpose of these two modifications is to increase its capture range and to obtain more realistic geometrical transformations between the surfaces. We performed some validation experiments and showed some results on synthetic and real data [36] .

A modified ICP algorithm for normal-guided surface registration

Participants : Daniel Münch, Benoît Combès, Sylvain Prima.

The ICP is probably the most popular algorithm for registration of surfaces. However, ICP-related registration methods suffer from the fact that they only consider the distance between the surfaces to register in the criterion to minimize, and thus are highly dependent on how the surfaces are aligned in the first place. This explains why these methods are likely to be trapped in local minima and to lead to erroneous solutions. A solution to partly alleviate this problem would consist in adding higher order information in the criterion to minimize (e.g. normals, curvatures, etc.), but previous works along these research tracks have led to computationally intractable minimization schemes. In this work, we proposed a new way to include the point normals in addition to the point coordinates to derive an ICP-like scheme for non-linear registration of surfaces and showed how to keep the properties of the original ICP algorithm with adequate implementation choices (most notably the use of a local, continuous, parametrization of the surfaces and a locally affine deformation model). We experimentally showed the strong added value of using the normals in a series of controlled experiments [37] .

Optimized supervised segmentation with Graph Cuts from multispectral MRIs

Participants : Jérémy Lecoeur, Christian Barillot.

We have proposed an optimized supervised segmentation method from multispectral MRIs. As MR images do not behave as natural images, using a spectral gradient based on a psycho-visual paradigm is sub-optimal. Therefore, we propose to create an optimized spectral gradient using multi-modalities MRIs. To that purpose, the algorithm learns the optimized parameters of the spectral gradient based on ground truth which are either phantoms or manual delineations of an expert. Using Dice Similarity Coefficient as a cost function for an optimization algorithm, we were able to compute an optimized gradient and to utilize it in order to segment MRIs with the same kind of modalities. Results show that the optimized gradient matrices perform significantly better segmentations and that the supervised learning of an optimized matrix is a good way to enhance the segmentation method. This has been applied to segment MS lesions with objective improved performances.

Supervised segmentation with Graph Cuts from multispectral images and atlas priors

Participants : Jérémy Lecoeur, Ryan Datteri, Christian Barillot.

The atlas based registration method uses both spatial and textural information, often resulting in a good segmentation. However, the search space is much too large to be comprehensively searched and, thus, some segmentations may have errors. The graph cut algorithm on the other hand is quick to compute and is able to use information from three separate image modalities, but it does not use any spatial information and can often be confused by organs that have a similar appearance. Also, the graph cut algorithm has the limitation of being semi-automatic. Therefore, the goal of this work was to combine both of these methods, creating better segmentations. A registration algorithm (affine or non-linear) is used to automate and initialize the graph cut algorithm as well as to add needed spatial information. Thanks to the multispectral implementation of the Graph Cut, the atlas prior is used as a complementary spatial information to multimodal observations in order to drive the segmentation to the most probable contours (from the observed images and from the probable location). Preliminary results on the segmentation of the Thalami shows better accordance than when using adapted atlas-based non-linear registration alone.

MAP Segmentation of 3D MR Images Based on Mean Shift and Markov Random Fields

Participants : Lei Lin, Christian Barillot.

In this work, we propose to combine mean-shift annealing, prior distribution coming from a probablistic atlas and Markov Random Field (MRF) to jointly estimate intensity inhomogeneities (to correct for bias field) and posterior maps of brain tissues. We employed the mean-shift algorithm to get a pixon-based image representation, and then the Markov random field (MRF) model was used to partition the image into a predefined number of tissue classes. The prior map coming from the SPM probabilistic template is then used to initialize the contribution of each tissue in individual pixon and a Bayesian framework is used to iterativelly estimate the global intensity inhomogeneity map plus the maximum a posteriori disctribution of each brain tissue class. The new method was validated on the simulated normal brain images from BrainWeb and on real brain images coming from IBSR. Compare with alternative MRI segmentation methods, the new method exhibited a higher degree of accuracy in segmenting real 3D MRI brain data.