Section: New Results

Image Segmentation, Registration and Analysis

Estimating A Reference Standard Segmentation with Spatially Varying Performance Parameters: Local MAP STAPLE

Participant : Olivier Commowick.

We present a new algorithm, called local MAP STAPLE, to estimate from a set of multi-label segmentations both a reference standard segmentation and spatially varying performance parameters. It is based on a sliding window technique to estimate the segmentation and the segmentation performance parameters for each input segmentation. In order to allow for optimal fusion from the small amount of data in each local region, and to account for the possibility of labels not being observed in a local region of some (or all) input segmentations, we introduce prior probabilities for the local performance parameters through a new maximum a posteriori formulation of STAPLE. Further, we propose an expression to compute confidence intervals in the estimated local performance parameters. We carried out several experiments with local MAP STAPLE to characterize its performance and value for local segmentation evaluation. First, with simulated segmentations with known reference standard segmentation and spatially varying performance, we show that local MAP STAPLE performs better than both STAPLE and majority voting. Then we present evaluations with data sets from clinical applications. These experiments demonstrate that spatial adaptivity in segmentation performance is an important property to capture. We compared the local MAP STAPLE segmentations to STAPLE, and to previously published fusion techniques and demonstrate the superiority of local MAP STAPLE over other state-of-the-art algorithms.

This work was done in collaboration with Alireza Akhondi-Asl and Simon K. Warfield [15] .

Voxel-based quantitative analysis of brain images from F-18 Fluorodeoxyglucose Positron Emission Tomography with a Block-Matching algorithm for spatial normalization

Participant : Olivier Commowick.

Statistical Parametric Mapping (SPM) is widely used for the quantitative analysis of brain images from F-18 fluorodeoxyglucose positron emission tomography (FDG PET). SPM requires an initial step of spatial normalization to align all images to a standard anatomic model (the template), but this may lead to image distortion and artifacts, especially in cases of marked brain abnormalities. This study aimed at assessing a block-matching (BM) normalization algorithm, where most transformations are not directly computed on the overall brain volume but through small blocks, a principle that is likely to minimize artifacts. Large and/or small hypometabolic areas were artificially simulated in initially normal FDG PET images to compare the results provided by statistical tests computed after either SPM or BM normalization. Results were enhanced by BM, compared with SPM, with regard to (i) errors in the estimation of large defects volumes (about 2-fold lower) because of a lower image distortion, and (ii) rates of false-positive foci when numerous or extended abnormalities were simulated. These observations were strengthened by analyses of FDG PET examinations from epileptic patients. Results obtained with the BM normalization of brain FDG PET appear more precise and robust than with SPM normalization, especially in cases of numerous or extended abnormalities.

This work was done in collaboration with Christophe Person, Valérie Louis-Dorr, Sylvain Poussier, Grégoire Malandain, Louis Maillard, Didier Wolf, Nicolas Gilet, Véronique Roch, Gilles Karcher and Pierre-Yves Marie [19] .

Block-matching strategies for rigid registration of multimodal medical images

Participants : Olivier Commowick, Sylvain Prima.

We propose and evaluate a new block-matching strategy for rigid-body registration of multimodal or multisequence medical images. The classical algorithm first matches points of both images by maximizing the iconic similarity of blocks of voxels around them, then estimates the rigid-body transformation best superposing these matched pairs of points, and iterates these two steps until convergence. In this formulation, only discrete translations are investigated in the block-matching step, which is likely to cause several problems, most notably a difficulty to tackle large rotations and to recover subvoxel transformations. We propose a solution to these two problems by replacing the original, computationally expensive, exhaustive search over translations by a more efficient optimization over rigid-body transformations. The optimal global transformation is then computed based on these local blockwise rigid-body transformations, and these two steps are iterated until convergence. We evaluate the accuracy, robustness, capture range and run time of this new block-matching algorithm on both synthetic and real MRI and PET data, demonstrating faster and better registration than the translation-based block-matching algorithm [28] .

Automated diffeomorphic registration of anatomical structures with rigid parts: Application to dynamic cervical MRI

Participants : Olivier Commowick, Sylvain Prima.

We propose an iterative two-step method to compute a diffeomorphic non-rigid transformation between images of anatomical structures with rigid parts, without any user intervention or prior knowledge on the image intensities. First we compute spatially sparse, locally optimal rigid transformations between the two images using a new block matching strategy and an efficient numerical optimiser (BOBYQA). Then we derive a dense, regularised velocity field based on these local transformations using matrix logarithms and M-smoothing. These two steps are iterated until convergence and the final diffeomorphic transformation is defined as the exponential of the accumulated velocity field. We show our algorithm to outperform the state-of-the-art log-domain diffeomorphic demons method on dynamic cervical MRI data [27] .

Computer-assisted paleoneurology

Participant : Sylvain Prima.

In collaboration with Antoine Balzeau and colleagues at the MNHN (http://www.mnhn.fr ), we made the first ever description of the “digital” endocranial cast of the Cro-Magnon 1 specimen, discovered in 1868 at the Eyzies-de-Tayac, Dordogne, France [13] . Together with Benoit Combès (Géosciences Rennes, UMR 6118), we were especially involved in the assessment of its endocranial asymmetries, using an algorithm previously developed at VisAGeS  [51] in the context of the ARC 3D-MORPHINE coordinated by Sylvain Prima (http://3dmorphine.inria.fr ).