Section: New Results
Image processing and modeling for biological imaging
Non-parametric regression for fluorescence microscopy image sequence denoising
Participants : Charles Kervrann, Patrick Bouthemy.
[In collaboration with J. Boulanger and P. Elbau (RICAM, Austria), J.-B. Sibarita and J. Salamero (Curie Institute)]
New video-microscopy technology enables to acquire 4-d data that require the design and the development of specific image denoising methods able to preserve details and discontinuities in both the (x-y-z ) space dimensions and the time dimension. Images are noisy due to the weakness of the fluorescence signal in time-lapse recording. Accordingly, in collaboration with UMR 144 CNRS / Curie Institute, we have developed an original and efficient spatio-temporal filtering method for significantly increasing the signal-to-noise ratio (snr) in noisy fluorescence microscopic image sequences where small particles have to be tracked from frame to frame. The proposed method exploits 3D+time information to improve the signal-to-noise ratio of images corrupted by mixed Poisson-Gaussian noise. A variance stabilization transform is first applied to the image-data to introduce independence between the mean and variance. This pre-processing requires the knowledge of parameters related to the acquisition system, also estimated in our approach. In a second step, we propose an original statistical patch-based framework for noise reduction and preservation of space-time discontinuities. In the continuous setting, we have shown that the proposed estimation procedure can be interpreted as a steepest descent algorithm related to the fixed point solution corresponding to the minimization of a global energy function involving non-local terms and local image contexts described by patches. The size of each neighborhood is optimized to improve the performance of the pointwise estimator. In our experiments, the snr is shown to be drastically improved and enhanced images can then be correctly segmented. In fluorescence video-microscopy, recent experiments demonstrated also that this method can be used as a pre-processing stage for image deconvolution, and allows us to increase the frame-rate of a factor 10 with the same snr values. Finally, this novel approach can be used for biological studies where dynamics have to be analyzed in molecular and subcellular bio-imaging.
Network tomography for tracking in fluorescence microscopy imaging
Participants : Thierry Pécot, Charles Kervrann, Patrick Bouthemy.
[In collaboration with J. Boulanger (RICAM, Austria), J.-B. Sibarita and J. Salamero (Curie Institute)]
The study of protein dynamics is essential for understanding the multi-molecular complexes at subcellular levels. Green Fluorescent Protein (GFP)-tagging and time-lapse fluorescence microscopy enable to observe molecular dynamics and interactions in live cells and consequently to make progress in knowledge about protein dynamics. Original image analysis methods are then required to process challenging 2D or 3D image sequences. Tracking methods that estimate the whole trajectories of moving objects have been successfully experimented but can be applied for tracking a limited number of objects (a few dozens). To address the tracking problem of several hundreds of objects, we propose instead an original framework that provides general information about molecule transport, that is about traffic flows between origin and destination regions detected in the image sequence. Traffic estimation can be accomplished by adapting the recent advances in Network Tomography commonly used in network communications. Our estimation method is inspired from the Network Tomography (NT) concept, introduced by Vardi in network communications and further applied to video surveillance. nt -based approaches, devoted to statistical traffic analysis, simplifies the tracking process because it only requires detection of an object as it moves from one region to another and avoids the difficult data association problem. This statistical method allows us to provide a global description of traffic flows. We just need to count the number of "objects/vesicles" in different image regions at each time step. In collaboration with UMR 144 CNRS / Curie Institute, we extended the usual NT concept to non-binary routing from geodesic paths given the image sequence. Unlike previous approaches, the new formulation can be considered as a probabilistic minimal paths modeling for object tracking. We showed that the origin-destination (OD)paths are not the minimal paths between the two extremities but formed as as a set of minimal paths joining intermediate points. We also we proposed an estimation/optimization framework to derive counting measurements from image intensity (fluorescence). The traffic flow problem is also solved with additional parsimonious constraints. This approach has been developed for real image sequences and Rab proteins, known to be involved in the regulation of intracellular membrane trafficking.
Repetitive and transient event detection in fluorescence video-microscopy
Participant : Charles Kervrann.
[In collaboration with J. Boulanger (RICAM, Austria), A. Gidon and J. Salamero (Curie Institute)]
Endocytosis-recycling is an essential cellular trafficking process regulating the proper distribution and function of a large set of molecules, such as lipids, receptors, or adhesion transmembrane proteins. This dynamic process also participates to the homeostasis of intracellular membrane compartments. Progresses in imaging dynamics behaviors of molecules including fast video microscopy and the application of evanescent wave microscopy have allowed to image intracellular vesicular movements, exocytosis and endocytosis of fluorescent-tagged proteins. In parallel, statistical image analysis has emerged as a basic methodology in the study of many biological phenomena. However, spatio-temporal analysis of transient events occurring at different sites of the cell has not been systematically performed. In addition, more formal tests are required in testing biological hypotheses, rather than visual inspection combined with more or less manual statistical analysis. For an unbiased quantification of repetitive and transient events, such as those observed during the trafficking of molecules traveling through the endosomal-recycling network of cells, their automatic detection become necessary. While requiring particular adjustments, our proposed approach is versatile enough, to be applicable to diverse although complementary modes of microscopy. This is illustrated for fast both video and TIRF imaging techniques in collaboration with UMR 144 CNRS / Curie Institute and RICAM (Austria). The proposed detection method can be decomposed into three main steps: i) a first pre-processing step is dedicated to the normalization of the image sequence; ii) the second step is the patch-based detection procedure to detect unusual patterns; iii) a third post-processing step allows us to cluster and count detected events in space and time. While focusing here on one particular Lectin receptor that constitutively recycles from internal compartments to the plasma membrane, it could be translated to many other studies of membrane trafficking in health and diseases such as diabetes, neurological, pigmentation or lysosomal defects.
Dynamic background subtraction in fluorescence video-microscopy
Participants : Charles Kervrann, Thierry Pécot, Patrick Bouthemy.
[In collaboration with A. Chessel, B. Cinquin and J. Salamero (Curie Institute)]
In collaboration with UMR 144 CNRS / Curie Institute, the main idea was to produce new descriptors able to capture spatio-temporal features of moving particles or group of particles with high variable motions. Actually, a wide range of proteins viewed using xFP probes in fluorescence video-microscopy shows a cytosolic state diffusing slowly and a membrane state eventually moving as vesicles or small tubules along the cell cytoskeleton network. Due to the property of fluorescence, the measured intensity is the sum of the contributions of these two components. The membrane and cytosolic components are then analyzed respectively by the Network Tomography approach for both the vesicle traffic estimation and the free-diffusion estimation in the cytosol. Previous parametric and non-parametric estimation methods have been tested for background subtraction. This year, we have proposed a promising and original framework based the computational geometry concept of -shapes. The relationships with the Empirical Mode Decomposition (EMD), the curvature motion-based scale-space and some operators from mathematical morphology, ware studied. We have also investigated the Conditional Random Field modeling framework to improve the background subtraction with more flexibility. This method has been compared to several algorithms : i) gray-scale opening (a.k.a. "rolling ball"); ii) wavelet-based detection combined with image interpolation; iii) Conditional Random Field-based detection; iv) computational geometry for temporal signal analysis. The evaluation protocol, related to the actual use of image sequence decomposition, includes a qualitative evaluation on real image sequences by experts and a quantitative evaluation on simulated sequences that mimic real images.
3D reconstruction in cryo-Electron Microscopy
Participants : Sophie Blestel, Charles Kervrann.
[In collaboration with D. Chrétien (UMR 6026, Rennes)]
In collaboration with UMR 6026 CNRS, we are interested in longitudinal projections of microtubules (i.e. perpendicular to the microtubule axis). Microtubules are composed of identical subunits that arrange together to form a helical lattice (see Figure 2). Because of their symmetry, most of their information can be retrieved from their Fourier spectrum. However, due to their flexibility, microtubules are generally curved and their Fourier transform can no longer be used for symmetry analysis.
We have proposed two sensitive contributions to automatically determine the local orientations and centers of short segments of microtubules in cryo-electron microscope images. Indeed, to our knowledge, the methods to determine the local centers of helices are not relevant for non centro-symmetric helices (e.g. 13-protofilament microtubules). The proposed algorithm exploits the helical symmetry of microtubules and the corresponding properties in the Fourier domain, so it can be easily extended to process other helical objects. Experimental results demonstrate that center locations are estimated with an accuracy of lower than one pixel. We have applied the algorithm to automatically straighten images of curved microtubules with odd numbers of protofilaments, and to improve 3D reconstructions of microtubules using back-projection methods.
Segmentation of microtubules in cryo-Electron Tomograms
Participants : Sophie Blestel, Charles Kervrann.
[In collaboration with D. Chrétien (UMR 6026, Rennes)]
Cryo-electron tomography allows 3D observation of biological specimens in their hydrated state. Generally, cryo-tomograms have very low signal-to-noise ratios, and conventional image segmentation methods yield poor results. To address this problem, we have considered the Conditional Random Fields (CRF) framework and we have formulated the segmentation problem as a maximum a posteriori estimation problem. Segmentation is obtained by computing the global minimum of a non-convex energy functional defined, in the discrete setting, as the sum of a fidelity term and of a regularization term. We define an original fidelity term robust to noise based on a distance between patches. Segmentation is performed section by section, with an automatic update of the reference patches for the 2 classes: `object' and `background'. Because of the contrast anisotropy in the specimen thickness direction, the whole tomogram is segmented section by section, with an automatic update of reference patches. This method has been is evaluated on synthetic data and on cryo-electron tomograms of in vitro microtubules.