Journal of

Structural Biology Journal of Structural Biology 144 (2003) 152–161 www.elsevier.com/locate/yjsbi

An improved algorithm for anisotropic nonlinear diffusion for denoising cryo-tomograms Jose-Jes us Fern andeza,* and Sam Lib a

Department of Computer Architecture and Electronics, University of Almerıa, 04120 Almerıa, Spain b MRC Laboratory of Molecular Biology, CB2 2QH Cambridge, UK Received 27 May 2003, and in revised form 5 August 2003

Abstract Cryo-electron tomography is an imaging technique with an unique potential for visualizing large complex biological specimens. It ensures preservation of the biological material but the resulting cryotomograms are extremely noisy. Sophisticated denoising techniques are thus essential for allowing the visualization and interpretation of the information contained in the cryotomograms. Here a software tool based on anisotropic nonlinear diffusion is described for filtering cryotomograms. The approach reduces local noise and meanwhile enhances both curvilinear and planar structures. In the program a novel solution of the partial differential equation has been implemented, which allows a reliable estimation of derivatives and, furthermore, reduces computation time and memory requirements. Several criteria have been included to automatically select the optimal stopping time. The behaviour of the denoising approach is tested for visualizing filamentous structures in cryotomograms.  2003 Elsevier Inc. All rights reserved. Keywords: Electron tomography; Denoising; Anisotropic nonlinear diffusion

1. Introduction Cryo-electron tomography (CryoET) is a useful tool for studying large biological specimens with a potential resolution of a few nanometers (Baumeister et al., 1999; Baumeister, 2002; Grunewald et al., 2003). Although it is not a new technique (Frank, 1992), the recent development of instruments and software has substantially changed its applicability (for a review of the technique, its evolution and prospects, see Frank, 1992; Koster et al., 1997; Baumeister et al., 1999; McEwen and Marko, 2001; Baumeister, 2002; Grunewald et al., 2003)). The current automation of low-dose data-acquisition procedures allows the study of 3D structures of unstained biological materials in near-physiological conditions. Therefore, it has potential to bridge the gap between cellular and molecular structural biology, and allows us to understand the cellular content in molecular details (Grunewald et al., 2003; Sali et al., 2003). Some examples of recent applications of CryoET to the anal* Corresponding author. Fax: +34-950-015-486. E-mail address: [email protected] (J.-J. Fernandez).

1047-8477/$ - see front matter  2003 Elsevier Inc. All rights reserved. doi:10.1016/j.jsb.2003.09.010

ysis of large pleiomorphic structures have been reported, including cellular organelles (Nicastro et al., 2000; Bohm et al., 2001; Nickell et al., 2003) and whole cells (Grimm et al., 1998; Medalia et al., 2002). Although great effort has been made for tomographic data collection, visualization of tomograms is cumbersome due to (1) low-dosage and low contrast imaging conditions, (2) limited number of projections along single or double tilt axes, and (3) large-size specimen embedded in thick ice (>300 nm). These result in poor signal-to-noise ratio (SNR) and severely hinder the interpretation of the tomogram. In the last instance, this low SNR precludes the direct application of image analysis techniques, such as automatic segmentation, pattern recognition, and techniques to extract useful information from the tomogram (Grunewald et al., 2003). Therefore it is evident that sophisticated filtering techniques are necessary before interpretation of the reconstructed tomogram (Frangakis and Hegerl, 2001; Frangakis et al., 2001; Medalia et al., 2002). Anisotropic nonlinear diffusion (AND) is a filtering method that reduces noise and enhances local structure which has been widely used in computational vision

J.-J. Fernandez, S. Li / Journal of Structural Biology 144 (2003) 152–161

153

(Weickert, 1998). It was first introduced in visualizing electron tomograms by Frangakis and Hegerl (2001) and Frangakis et al. (2001) and shows a superior performance to other filtering methods. Here, we report the development of a software program for filtering cryotomograms based on AND, which has additional improved functionalities suitable for visualizing 3D tomograms from single tilt-axis CryoET. Fig. 1. Basic local structures. l1 ; l2 ; l3 are the eigenvalues of the structure tensor. v1 ; v2 ; v3 are the corresponding eigenvectors.

2. Overview of anisotropic nonlinear diffusion AND can be considered as an adaptive gaussian filtering technique in which, for every voxel in the volume, an anisotropic 3D gaussian function is computed whose widths and orientations depend on the local structure (Barash, 2002). This section presents local structure determination via structure tensors, the concept of diffusion and, finally, three common diffusion approaches. 2.1. Description of local structure by structure tensors Local structure in a multidimensional image can be estimated by means of the so-called structure tensor (also known as second moment tensor). Let IðxÞ denote a 3D image, where x ¼ ðx; y; zÞ is the coordinate vector. The structure tensor of the image is a symmetric positive semidefinite matrix given by 2 2 3 Ix Ix Iy Ix Iz T 2 J0 ðrIÞ ¼ rI  rI ¼ 4 Ix Iy Iy Iy Iz 5; ð1Þ 2 Ix Iz Iy Iz Iz where Ix ¼ oI=ox, Iy ¼ oI=oy, Iz ¼ oI=oz are the derivatives of the image density with respect to x, y and z, respectively. Under some conditions (see below), it is more suitable to work with the averaged structure tensor, which takes the form Jr ðrIÞ ¼ Kr  ðrI  rI T Þ;

ð2Þ

where the symbol ‘‘’’ stands for componentwise convolution, and Kr is a gaussian convolution kernel, with r being the integration scale over which the local structure is averaged. The eigen analysis of the structure tensor allows characterization of the local structural features in the image at a Kr scale (Weickert, 1998). The orthogonal eigenvectors v1 ; v2 ; v3 provide the preferred local orientations, and the corresponding eigenvalues l1 ; l2 ; l3 (assume l1 P l2 P l3 ) provide the average contrast along these directions. The first eigenvector v1 represents the direction of the maximum variance, whereas v3 points to the direction with the minimum variance. The three eigenvalues could therefore be used, based on their relative values, to describe local structures in three

classes: line-like or plane-like or isotropic structure, as illustrated in Fig. 1: • Line-like structures: Characterized by a preferred direction (v3 ) exhibiting a minimum variation whose eigenvalue is much lower than the other two. v1 and v2 represent directions perpendicular to the line. • Plane-like structures: There are two preferred directions exhibiting similar small contrast variation, whose eigenvalues are much lower than the first one. v1 represents the direction perpendicular to the plane-like structure, while v2 and v3 define the plane. • Isotropic structures: Exhibit eigenvalues of similar magnitude. 2.2. Diffusion Diffusion is a physical process that equilibrates concentration differences as a function of time, without creating or destroying mass. This is expressed by the diffusion equation (Weickert, 1998): It ¼ divðD  rIÞ;

ð3Þ

where It ¼ oI=ot denotes the derivative of the image density I with respect to the time t, rI is the gradient vector, D is a symmetric square matrix called diffusion tensor and div is the divergence operator. The divergence is defined for a vector function f ¼ ðfx ; fy ; fz Þ as divðfÞ ¼

ofx ofy ofz þ þ : ox oy oz

In AND the smoothing depends on both the strength of the gradient and its direction measured at a local scale. The diffusion tensor D is therefore defined as a function of the structure tensor J0 or its averaged version Jr : 2 3 k1 0 0 T ð4Þ D ¼ ½v1 v2 v3   4 0 k2 0 5  ½v1 v2 v3  ; 0 0 k3 where vi denotes the eigenvectors of the structure tensor. The eigenvalues ki of the diffusion tensor D are computed as a function of the eigenvalues li of the structure tensor. This approach applies diffusion anisotropically according to the eigendirections of the local structure of

154

J.-J. Fernandez, S. Li / Journal of Structural Biology 144 (2003) 152–161

the image. The strength of the diffusion along them is based on the corresponding eigenvalues ki of D.

applied. This approach reduces the background noise and enhances flow-like structures. It is an efficient approach for visualizing tomograms.

2.3. Diffusion approaches Depending on ki in D, AND acts in several ways, either filtering noise or enhancing some structural features in the volume. Currently, the most commonly used diffusion approaches are EED and CED. Frangakis and Hegerl (2001) proposed a CED/EED Hybrid approach. dEED: edge enhancing diffusion. The primary effects of EED are edge preservation and edge enhancing (Weickert, 1998). In EED, the diffusion tensor D is computed from the non-averaged structure tensor J0 , and ki are set up as 8 < k1 ¼ gðjrIjÞ; k ¼ gðjrIjÞ; ð5Þ : 2 k3 ¼ 1; with g being a monotonically decreasing function, such as (Weickert, 1998) ! 3:31488 gðjrIjÞ ¼ 1  exp ; ðjrIj=KÞ8 where K > 0 is a contrast threshold constant; Structures with jrIj > K are regarded as edges, otherwise as the interior of a region. With these ki , a strong smoothing is carried out along the direction of the minimum change (v3 ). The diffusion strength along v1 and v2 depends inversely on the gradient. Therefore, the effect of smoothing is carried out along the edges, instead of across them, and hence edges are preserved and enhanced. dCED: coherence enhancing diffusion. CED focuses on the coherence of the curvilinear continuity and is intended to improve flow-like structures (Weickert, 1999a). The diffusion tensor D is computed from the averaged structure tensor Jr and ki are set up as 8 > < k1 ¼ a; k2 ¼ a; ð6Þ   > C : k3 ¼ a þ ð1  aÞ exp ; 2 ðl l Þ 1

3

where a is a regularization constant (typically 103 ) and C > 0 is a threshold. Structures with ðl1  l3 Þ2 > C will be regarded as line-like patterns. These ki make smoothing along v3 . Since it points to the major direction along the structure, line-like structures are thus enhanced. dCED/EED hybrid approach. The original method proposed by (Frangakis and Hegerl, 2001) and (Frangakis et al., 2001) provided a hybrid approach combining the advantages of EED and CED. It is based on the fact that the difference ðl1  l3 Þ represents the local relation of structure and noise. Therefore, it can be used as a switch: CED is applied if this value is larger than a suitable threshold parameter, otherwise EED is

3. Plane-like coherence enhancing diffusion and the hybrid approach The program presented here is also based on the CED/EED hybrid approach. In addition, the CED strategy is used to enhance plane-like (or surface-like) structures as well as line-like structures. 3.1. Enhancement of plane-like structures The standard CED approach diffuses unidimensionally along the third eigendirection v3 , which is the direction of the minimum change. Therefore, it is efficient for enhancing line-like structures (where l1  l2  l3 ). However, in biological specimens, a significant amount of structural features are present as plane-like at local scales: membranes, surfaces of organelles, even compact macromolecules where their surfaces can be considered as curved planes. In general, any surface of an object in a cryotomogram can be considered as a plane-like structure at a local scale. In such structures, there are two eigendirections v2 and v3 with similar small contrast variation (l1  l2  l3 ). For enhancing planar structures, multidimensional CED diffusion along both v2 and v3 is thus necessary (Weickert, 1999a). In addition, the resolution in single-axis tomography is strongly affected by the missing wedge (Frank, 1992). The effect of the missing wedge on the structure tensor was tested on a model of a microtubule. Microtubules are tubular structures. At local scales all the voxels in the surface are considered as part of a plane-like structure. A microtubule model at 4 nm/voxel was computed from the PDB atomic coordinates. A tilt series of 61 projections in ½60; 60 at a 2 interval was computed from the model, with the tilt axis perpendicular to the tubular axis. These projections were used to reconstruct the tomogram by weighted backprojection. Fig. 2 shows the model and the reconstruction. The reconstructed

Fig. 2. Microtubule model and its single-axis reconstruction. The tilt axis was located perpendicular to the microtubule helical axis. Sixtyone projections in tilt range ½60; 60 at intervals of 2 were used.

J.-J. Fernandez, S. Li / Journal of Structural Biology 144 (2003) 152–161

microtubule appears as two parallel walls instead of a hollow cylindrical structure. To quantify the effect of the missing wedge, the angle between the Z-axis and the eigenvectors was computed for all the surface-like voxels in the model and the reconstruction. The range of possible values for the angle between the Z-axis and the eigenvectors (½0; 90) was discretized into nine bins (see Fig. 3A). The first bin represents the range of ½0; 10 with respect to the Z-axis, which is the bin closest to the direction of the Zaxis. The last bin represents the range ½80; 90, i.e. the closest to the XY plane and the farthest from the Z-axis. Fig. 3B shows the histograms (in %) of the distribution of the angle between Z-axis and the three eigenvectors for the microtubule model and the reconstruction. Abscissa represents the angle with respect to the Z-axis, discretized into the nine bins. In the original model, v1 , which is perpendicular to the microtubule surfaces, is distributed throughout the range nearly uniformly. v3 is clearly located on the XY plane, parallel to the tubular axis. However, the histograms of the single-axis reconstruction change substantially. v1 is instead predominantly located on the XY plane, which indicates that the reconstruction is made up of walls or planes perpendicular to the XY plane. v3 is in turn distributed throughout the angle range. These results show that, due to the missing wedge, structure tensors in the tomogram have a tendency with the first eigendirection v1 going downwards (towards the plane XY ) and v3 upwards (towards Z). The second eigenvector is also affected, but the new direction is not so relevant since it is given by the other eigenvectors, as all of them must be orthogonal. In this work, other models (a hollow sphere and planes in different directions) and other orientations of the single tilt axis have been tested. Similar effects of the missing wedge have been observed (data not shown).

155

The tests that were carried out indicate that, in singleaxis tomography, the eigendirections measured from the tomograms are severely distorted compared to the directions in the original specimen. In the case of planelike structures, if the diffusion were restricted along the third eigenvector v3 , it would then be driven only by one of the vectors defining the plane. Moreover, v3 may misrepresent the feature because of the distortion. In these cases, the diffusion should involve v2 and v3 . Although distorted, the combined information provided by both is valuable to find out the planar information. Therefore, diffusion using second and third eigendirections is essential for successfully denoising and enhancing plane-like structures, and the improvement is more significant as stronger effects due to the missing wedge is present in the reconstruction. 3.2. Detection of plane-like structures In practice, a reconstructed tomogram contains both plane-like and line-like structures. Therefore, it is necessary to develop a mechanism to discern these features and meanwhile, to avoid artifacts when applying CED. In this work, we have defined a set of metrics to estimate whether the features are plane-like, line-like or isotropic. Let P1 , P2 , and P3 be the following metrics: 2 P1 ¼ l1ll 1 l2 l3 P2 ¼ l1 P3 ¼ ll31 which satisfy 0 6 Pi 6 1, 8i and P1 þ P2 þ P3 ¼ 1, and where l1 , l2 , and l3 are the eigenvalues of the averaged structure tensor Jr . These metrics P1 , P2 , and P3 provide, respectively, measures about the planar, linear or isotropic nature of the local structure. A voxel is said to be part of one of those structures according to the following conditions:

Fig. 3. (A) The angle between the Z-axis and the eigenvectors is discretized into 9 bins. (B) Histograms of the angle between Z-axis and the three eigenvectors of the microtubule model. On the left, the histograms for the original model. On the right, the histogram for the single-axis reconstruction.

156

J.-J. Fernandez, S. Li / Journal of Structural Biology 144 (2003) 152–161

P1 > P2 and P1 > P3 ) plane-like; P2 > P1 and P2 > P3 ) line-like; P3 > P1 and P3 > P2 ) isotropic: These metrics allow the diffusion method to proceed safely without producing artifacts in the structures. CED along second and third eigendirections is only applied for voxels classified as part of plane-like structures. 3.3. The final diffusion approach The following is the outline of our approach for AND: 1. Determination EED vs CED. This step is intended to determine if the voxel is to be processed as EED or CED. The voxel is considered CED if the local relation of structure and noise, given by l1  l3 from J0 , is larger than a suitable threshold. Otherwise, the voxel is considered EED. The threshold is computed from the mean value of ðl1  l3 Þ in a subvolume of the image containing only noise. This subvolume is previously delimited by the user. 2. EED: edge enhancing diffusion. If the voxel is classified as EED, the diffusion tensor D (see Eq. (4)) is computed from the non-averaged structure tensor J0 , and ki are set up according to Eq. (5). 3. CED: coherence enhancing diffusion. If the voxel is classified as CED, the steps to follow are: 3.1. Determination if planar local structure. The voxel is considered part of a plane-like structure if the conditions in Section 3.2 are met. 3.2. Planar enhancing diffusion. If it does belong to a plane, the diffusion tensor D is computed from the averaged structure tensor Jr , and ki are set up to allow diffusion along the second and third eigendirections: 8 k1 ¼ a; >   > < 2 k2 ¼ a þ ð1  aÞ exp ðl C ; 2  1 l2 Þ  > > C3 : k3 ¼ a þ ð1  aÞ exp : ðl l Þ2 1

The diffusion equation (Eq. (3)) can be numerically solved using finite differences. The term It ¼ oI=ot can be replaced by an Euler forward difference approximation. The resulting explicit scheme allows calculation of the image at a new time step directly from the version at the previous step:  o o o ðD11 Ix Þ þ ðD12 Iy Þ þ ðD13 Iz Þ I ðkþ1Þ ¼ I ðkÞ þ s  ox ox ox o o o þ ðD21 Ix Þ þ ðD22 Iy Þ þ ðD23 Iz Þ oy oy oy  o o o þ ðD31 Ix Þ þ ðD32 Iy Þ þ ðD33 Iz Þ ; ð7Þ oz oz oz where s denotes the time step size, I ðkÞ denotes the image at time tk ¼ ks and Ix ¼ oI=ox, Iy ¼ oI=oy, and Iz ¼ oI=oz are the derivatives of the image density with respect to x, y, and z, respectively. The Dmn terms represent the components of the diffusion tensor D. The standard explicit numerical scheme (Weickert, 1998) for solving the partial differential equation (PDE) in Eq. (7) is based on central differences to approximate the spatial derivatives (o=ox, o=oy, and o=oz). The standard approach then involves a 3  3  3-stencil in solving the PDE. In this work, the spatial derivatives have been approximated by derivative filters that have been proved to be optimal (Jahne et al., 1999). These filters approximate rotational invariance significantly better than traditional kernels (Weickert and Scharr, 2002). In order to provide the AND approach with better capabilities for structural preservation, a 3D version of these rotational-invariant filters has been derived (see Appendix A for details of the derivation). The use of these 3  3  3 filters results in a more reliable approximation of the derivatives. Furthermore, it involves a 5  5  5-stencil in the solution of the PDE. Finally, the rotational invariance allows preservation of finer details because the gradients account for curvatures in the structures (Weickert and Scharr, 2002).

3

3.3. Curvilinear enhancing diffusion. Otherwise, the standard CED is applied. D is computed from Jr , and ki are set up to diffuse only along the third eigendirection according to Eq. (6). 4. Implementation 4.1. Discretization scheme In this work, a novel discretization scheme is proposed for AND in three dimensions. It is based on Euler forward explicit numerical schemes and uses derivative filters to approximate the spatial derivatives in the diffusion formulae.

4.2. Efficiency of the discretization scheme The explicit numerical scheme based on 3D rotationally invariant kernels for the spatial derivatives has been proved (Weickert and Scharr, 2002) to allow a four times larger time step size (s ¼ 0:4) than the traditional explicit scheme (s ¼ 0:1). This is due to the use of larger stencils in the PDE solution, which makes the discretization scheme more stable. Consequently, the program presented here requires 4 times less iterations than the traditional scheme. However, the scheme based on a 5  5  5-stencil is more computationally complex and every iteration takes more time than the traditional scheme. At the end, our implementation exhibits a net speedup of the AND

J.-J. Fernandez, S. Li / Journal of Structural Biology 144 (2003) 152–161

method by about 1.5–3.0 compared to the traditional scheme. 4.3. Memory requirements Denoising based on AND combining the EED and CED approaches has huge memory requirements. Structure tensors, input and output volumes, and some other additional matrices are needed, which makes the method require up to 15 times the size of the input volume. Our implementation has optimized memory usage. It uses only one copy of the matrix for the structure tensors. Since a given voxel is processed either as EED or CED, it is possible to combine both tensors J0 ðrIÞ and Jr ðrIÞ in the same matrix. However, an additional matrix which only requires one bit per voxel is then needed to indicate if a voxel is to be processed as EED or CED. This matrix has a negligible size compared to the input volume. Furthermore, the gradient modulus jrIj required for EED voxels is generated on-the-fly from the combined structure tensor (see Eq. (1)) since jrIj ¼ Ix2 þ Iy2 þ Iz2 . Consequently, memory requirements here are eight times the size of the input volume: two for the volumes in the current and previous iterations and six for the structure tensors. An additional relatively small bit-matrix storing the CED/EED information is also required. It would be possible to further reduce the memory requirements for the structure tensors to a minimum at expenses of computation time. By using a sliding window of 5  5  5, the structure tensors could also be computed on-the-fly. However, the penalty in computation time might be significant. 4.4. The algorithm of the diffusion approach The algorithm for solving the PDE in Eq. (7) using the discretization of the temporal and spatial derivatives described above consists of the following steps: 1. Compute the structure tensor combining J0 and Jr . 2. Compute the diffusion tensor D from the corresponding entries of J0 or Jr , according to the strategy described in Section 3.3. 3. Compute the resulting image at the current step k from the previous step k  1 by means of Eq. (7). The resulting image corresponds to the diffusion time tk ¼ ks. This algorithm is executed iteratively for a number of iterations N . The final image is the result after a total diffusion time T ¼ tN ¼ N s.

5. Stopping criterion AND is an iterative denoising method which produces successive filtered versions of the image (see Eq.

157

(7)). A crucial question is when to stop the filtering process, so that the signal in the image is not significantly affected by the denoising. Several objective stopping criteria have been proposed (see (Mrazek and Navara, 2003) for a brief review). In this work, some of them have been implemented and tested. The first stopping approach (Weickert, 1999b) is based on the fact that the relative variance (ratio between the variances of the filtered image at time t, I t , and the original noisy image I 0 ) decreases monotonically from 1 to 0 during diffusion. A stopping criterion may be defined by a threshold over the relative variance. A suitable threshold is also proposed based on a-priori knowledge on the variance of the noise (Weickert, 1999b). However, in most cases this threshold underestimates the optimal stopping time (Mrazek and Navara, 2003). Therefore, a threshold based on the desired reduction factor of the variance with respect to that of the original image is more appropriate. For instance, a threshold of 0.4 would mean that the denoising process should stop when the image I t exhibits less than 0:4 the variance of the input image I 0 . The formula of relative variance is given by rðtÞ ¼

varðI t Þ : varðI 0 Þ

The second criterion, Decorrelation criterion (Mrazek and Navara, 2003), assumes that signal and noise are uncorrelated. It is based on the correlation between the filtered image at time t, I t , and the noise. The noise is estimated as the difference between the original noisy image I 0 and the current filtered version, I t . This correlation should decrease, meaning that the noise that was eliminated and the signal are not correlated. However, if that correlation increases, which indicates the signal is starting being affected, then the diffusion has to be stopped. This criterion defines optimal stopping time at minimum of the correlation coefficient between the estimates of the noise and signal tstop ¼ arg mint jcorrðI 0  I t ; I t Þj: The decorrelation criterion estimates the optimal diffusion stopping time without any a-priori knowledge on the signal and noise. However, this criterion is not guaranteed to be unimodal nor exhibits a single minimum (Mrazek and Navara, 2003). In some applications the first local minimum coincides with the global one and thus the criterion is still valid (Mrazek and Navara, 2003). Our experience of using this criterion is that, in general, it exhibits non-unimodal tendencies and thus is not reliable for defining a suitable time to stop the filtering. Finally, we propose a stopping criterion based on the evolution of the variance in the subvolume of noise from which the threshold for the EED/CED switch is also computed (see Section 3.3). The relative noise variance is

158

J.-J. Fernandez, S. Li / Journal of Structural Biology 144 (2003) 152–161

the ratio between the variance of the noise subvolume at time t, INt , and the variance of the original subvolume, IN0 :   var INt rN ðtÞ ¼ : varðIN0 Þ This curve decreases monotonically from 1 to 0 and, as for the relative variance. A stopping criterion may then be established by defining a suitable threshold based on the desired noise reduction factor. For example, a threshold of 0:1 would mean that the denoising process should stop when the remaining variance in the noise subvolume INt is less than 0:1 the variance of the subvolume in the input image IN0 . 6. Application As an illustration, the results from the application of AND to tomograms of microtubules embedded in ice are presented and discussed. The tomogram was reconstructed through weighted backprojection of a tilt series from ½60; 60, at 1.5 interval. The images were recorded with an FEI Tecnai F30 TEM at 300 kV. The reconstructed tomogram was finally downsampled to 4.0 nm/voxel. The tomogram was filtered using AND with the features described in the previous sections. In order to compare the plane-enhancing CED described in Section 3, the denoising was carried out using EED combined with (1) plane-enhancing CED (diffusion along v2 and v3 ) according to Section 3.3 and (2) only curvilinear-enhancing CED (diffusion along v3 ). In both cases, the parameters for K for EED, and C for CED were the same. Fig. 4 shows the curves of the three stopping criteria tested here. On the left, the curves of the relative variance (RV) and the relative noise variance (RNV) are shown. On the right, the curves for the correlation coefficient between the estimates of noise and signal in the

decorrelation criterion are shown. RNV exhibits the same curve for both cases and thus they overlap. According to RV, plane-enhancing CED is clearly better since from the beginning the variance decrease is larger. With regard to the decorrelation criterion on the right, the correlation curve exhibits a single minimum for curvilinear-enhancing CED, whilst its tendency is far from unimodal for plane-enhancing CED. The AND denoising process was stopped at the seventh iteration, equivalent to a time step of t ¼ 2:8, where RNV reaches 0.1. RV in turn exhibits at that iteration a value below 0.4 in the case of plane-enhancing CED, whereas curvilinear-enhancing CED needs 25 iterations to reach similar values of RV. The result of curvilinearenhancing CED at 25 iterations (not shown here) exhibits a more homogeneous background, but no significant improvement in the CED areas, i.e. the reduction of variance is mainly due to noise reduction. According to the decorrelation criterion, the diffusion should also stop at the seventh iteration in the case of curvilinear-enhancing CED, since the minimum of the curve is located there. However, in the case of planeenhancing CED, the correlation does not exhibit a curve with a well defined stopping time. Nevertheless, it seems that around the seventh iteration this curve shows a substantial change of slope that might be considered as an indication. Fig. 5 shows the denoising results for a slice of the tomogram. The noise filtering with respect to the original slice is evident in both denoising cases. However, the enhancing and smoothing of the microtubules are more apparent in the case of plane-enhancing CED. The continuity of the microtubules is now significant. Not only do the microtubules look much better, but the area just around them is also further smoothed, allowing a better definition of their boundaries. However, the result from curvilinear-enhancing CED does not exhibit a significant enhancement of the microtubules with respect to the original slice.

Fig. 4. Convergence curves corresponding to the different stopping criteria. Left: relative variance and relative noise variance criteria. Right: decorrelation criterion.

J.-J. Fernandez, S. Li / Journal of Structural Biology 144 (2003) 152–161

159

Fig. 6. Isosurface representation of a part of amicrotubule. From top to bottom: the original microtubule area; denoised version with EED combined with plane-enhancing CED; denoised version with EED combined with curvilinear-enhancing.

walls, without disruptions. However, the result from curvilinear-enhancing CED is not so continuous, and still contains some disruptions and small artifacts along the microtubule.

7. Discussion and conclusion

Fig. 5. Denoising with AND using EED and CED. From top to bottom: the original slice; denoised version with EED combined with plane-enhancing CED; denoised version with EED combined with curvilinear-enhancing CED. K for EED, C for CED and time step were set up to 0.018, 5  108 , and 2.8, respectively. The visualized density range is l  4  r, where l and r are the mean and standard deviation of the corresponding tomogram. The tilt axis is along the vertical direction.

Fig. 6 shows the results of the isosurface representation of a part of a microtubule in the tomogram. Clearly, the original tomogram looks very noisy, and denoising significantly improves the visualization of the microtubule. Both denoising approaches yield results much cleaner than the original. The effect of the missing wedge is evident in the pictures. The microtubule looks like a couple of parallel walls instead of a tubular structure. The microtubule resulting from the planeenhancing CED exhibits more continuity along the

Filtering techniques are necessary to allow visualization and interpretation of the information about the supramolecular organization of biological specimens obtained from CryoET. In this paper a software tool for filtering cryotomograms based on AND has been described. The approach is based on the hybrid EED/CED denoising approach first introduced by Frangakis and Hegerl (2001). The program described here contains a number of improvements, including recent advances in computer vision. First, the program includes a new diffusion mode, plane-enhancing CED, which enhances surface-like or plane-like local structures in tomograms. This CED mode turns out to be useful in general, as tomograms normally exhibit a significant amount of such structural features. This work has shown a quantitative analysis of the effects of the missing wedge in single-axis tomography on the distorting change in the eigenvectors of the local structures. The use of plane-enhancing CED allows a better enhancement of the distorted surface- or planelike local structures. This improvement has been shown

160

J.-J. Fernandez, S. Li / Journal of Structural Biology 144 (2003) 152–161

for the particular example of cryotomograms of microtubules, where the distorting effect of the missing wedge may be particularly dramatic. The software tool uses a novel discretization scheme that has been derived to solve the partial differential equation in the diffusion process. Compared to the standard numerical scheme, this one is more reliable for computing derivatives and exhibits better rotational invariance. One of the direct consequences of this scheme is the stability in the PDE solution. This results in a reduction of the number of iterations by a factor of 4, and a net speedup by about 1.5–3.0, depending on the computer. The program also reduces the memory requirements. Previously, the hybrid EED/CED approach required up to 15 times the size of the input tomogram. The present tool needs eight times the size of the tomogram plus a relatively small auxiliary matrix. As an example of the memory needs, denoising a tomogram of 100 Mbytes in size would require around 803 MB, compared to about 1.5 GB previously. Finally, the program provides a set of stopping criteria. The relative variance and relative noise variance seem to be the most useful ones. The former takes into account the entire tomogram to compute the metric, whereas the latter considers only a noise subvolume. By defining a threshold over those metrics, it is possible to find an optimal stopping time of the diffusion process. The thresholds should be set according to the desired variance reduction. In this work, it was seen that thresholds around 0.4 for the relative variance and 0.1 for the relative noise variance are reasonable. The decorrelation criterion turns out to be of limited applicability. It is based on the unimodality of the curve, but the curves from the tomograms in this work have proved to be, in general, monotonically decreasing. Consequently, it has been difficult to apply this criterion in practice. This could be due to the combined use of EED and CED. The criterion was introduced for linear and nonlinear diffusion and might not be suitable for EED/CED hybrid anisotropic diffusion. Further work is needed to stablish an objective criterion which takes into consideration both the variance reduction and the signal enhancement for estimating a suitable stopping diffusion time. Such a criterion would stop the diffusion in the time step when the signal starts being affected, for example, by too much CED, thus smearing out interesting structural features.

Acknowledgments The author wishes to thank Drs. Frangakis and Hegerl for kindly providing the code of AND, Dr. R.A. Crowther for critical reading the manuscript and their support and help throughout the work, and Drs. R. Henderson, A.M. Roseman and P.B. Rosenthal for

helpful discussions. This work has been partially developed while JJF was visiting the MRC Laboratory of Molecular Biology. SL is supported by a HFSP longterm fellowship. This work has been supported by the MRC (UK) and the Spanish CICYT grant TIC200200228.

Appendix A. Derivation of kernels for computing spatial derivatives with optimally directional invariance The spatial derivatives in the diffusion formulae have been approximated by derivative filters with optimally directional invariance. The 3D version of the filter has been derived by using the techniques described in (Jahne et al., 1999). In essence, the filter is derived by the convolution of a 1D derivative kernel (½1; 0; 1=2, denoted by D) and 1D smoothing kernels (denoted by B) in all other directions. The optimal value of B for a 3  3  3 derivative filter was found to be: [0.174654, 0.650692, 0.174654] (see (Jahne et al., 1999) for details). The kernels to approximate the derivatives are thus computed as o ¼ Dx  B y  B z ; ox o ¼ Dy  Bz  Bx ; oy o ¼ Dz  B x  B y ; oz where the subscripts x; y; z in D and B represent the direction of the kernel, and  is a convolution. That way, the 3  3  3 kernel used for computing the derivative o=ox (the kernels for o=oy and o=oz are computed similarly) results in 3 32 2 32 0:0305 0 0:0305 0:1136 0 0:1136 0:0305 0 0:0305 14 0:1136 0 0:1136 54 0:4234 0 0:4234 54 0:1136 0 0:1136 5: 2 0:0305 0 0:0305 0:1136 0 0:1136 0:0305 0 0:0305

References Barash, D., 2002. A fundamental relationship between bilateral filtering, adaptive smoothing and the nonlinear diffusion equation. IEEE Trans. Pattern Anal. Mach. Intel. 24, 844–847. Baumeister, W., 2002. Electron tomography: towards visualizing the molecular organization of the cytoplasm. Curr. Opin. Struct. Biol. 12, 679–684. Baumeister, W., Grimm, R., Walz, J., 1999. Electron tomography of molecules and cells. Trends Cell Biol. 9, 81–85. Bohm, J., Lambert, O., Frangakis, A., Letellier, L., Baumeister, W., Rigaud, J., 2001. FhuA-mediated phage genome transfer into liposomes. A cryo-electron tomography study. Curr. Biol. 11, 1168–1175. Frangakis, A., Hegerl, R., 2001. Noise reduction in electron tomographic reconstructions using nonlinear anisotropic diffusion. J. Struct. Biol. 135, 239–250.

J.-J. Fernandez, S. Li / Journal of Structural Biology 144 (2003) 152–161 Frangakis, A., Stoschek, A., Hegerl, R., 2001. Wavelet transform filtering and nonlinear anisotropic diffusion assessed for signal reconstruction performance on multidimensional biomedical data. IEEE Trans. Biomed. Eng. 48, 213–222. Frank, J. (Ed.), 1992. Electron Tomography. Three-Dimensional Imaging with the Transmission Electron Microscope. Plenum Press, New York. Grimm, R., Singh, H., Rachel, R., Typke, D., Zillig, W., Baumeister, W., 1998. Electron tomography of ice-embedded prokaryotic cells. Biophys. J. 74, 1031–1042. Grunewald, K., Medalia, O., Gross, A., Steven, A., Baumeister, W., 2003. Prospects of electron cryotomography to visualize macromolecular complexes inside cellular compartments: implications of crowding. Biophys. Chem. 100, 577–591. Jahne, B., Scharr, H., Korkel, S., 1999. In: Jhne, B., Hauecker, H., Geile, P. (Eds.), Handbook of Computer Vision and Applications, Vol. 2: Signal Processing and Pattern Recognition. Academic Press, San Diego, pp. 125–152 (Chapter: Principles of Filter Design). Koster, A., Grimm, R., Typke, D., Hegerl, R., Stoschek, A., Walz, J., Baumeister, W., 1997. Perspectives of molecular and cellular electron tomography. J. Struct. Biol. 120, 276–308. McEwen, B., Marko, M., 2001. The emergence of electron tomography as an important tool for investigating cellular ultrastructure. J. Histochem. Cytochem. 49, 553–564.

161

Medalia, O., Weber, I., Frangakis, A., Nicastro, D., Gerisch, G., Baumeister, W., 2002. Macromolecular architecture in eukaryotic cells visualized by cryoelectron tomography. Science 298, 1209– 1213. Mrazek, P., Navara, M., 2003. Selection of optimal stopping time for nonlinear diffusion filtering. Int. J. Comput. Vision 52, 189–203. Nicastro, D., Frangakis, A., Typke, D., Baumeister, W., 2000. Cryoelectron tomography of neurospora mitochondria. J. Struct. Biol. 129, 48–56. Nickell, S., Hegerl, R., Baumeister, W., Rachel, R., 2003. Pyrodictium cannulae enter the periplasmic space but do not enter the cytoplasm, as revealed by cryo-electron tomography. J. Struct. Biol. 141, 34–42. Sali, A., Glaeser, R., Earnest, T., Baumeister, W., 2003. From words to literature in structural proteomics. Nature 422, 216–225. Weickert, J., 1998. Anisotropic Diffusion in Image Processing, Teubner. Weickert, J., 1999a. Coherence-enhancing diffusion filtering. Int. J. Comput. Vision 31, 111–127. Weickert, J., 1999b. Coherence-enhancing diffusion of colour images. Image Vision Comput. 17, 201–212. Weickert, J., Scharr, H., 2002. A scheme for coherence-enhancing diffusion filtering with optimized rotation invariance. J. Visual Commun. Image. Represent. 13, 103–118.

An improved algorithm for anisotropic nonlinear ...

development of a software program for filtering cryot- omograms based on AND, ... AND can be considered as an adaptive gaussian filtering technique in which, ...

364KB Sizes 1 Downloads 328 Views

Recommend Documents

An improved Incremental Delaunay Triangulation Algorithm
Abstract: The incremental insertion algorithm of. Delaunay triangulation in 2D is very popular due to its simplicity and stability. We briefly discuss.

An Improved Divide-and-Conquer Algorithm for Finding ...
Zhao et al. [24] proved that the approximation ratio is. 2 − 3/k for an odd k and 2 − (3k − 4)/(k2 − k) for an even k, if we compute a k-way cut of the graph by iteratively finding and deleting minimum 3-way cuts in the graph. Xiao et al. [23

An Improved Acquisition Algorithm for the Uplink ...
CDMA (Code Division Multiple Access) wireless communications system is a .... on the acquisition receiver invented in [4]. The accumulation of the signal ...

An Improved Acquisition Algorithm for the Uplink ...
and the error floor of the traffic transmission caused by imperfect synchronization can be .... The length of a frame is 20 ms, and the system load is 4 or 8 users.

An Improved Algorithm for the Solution of the ...
tine that deals with duplicate data points, a routine that guards against ... numerical algorithm has been tested on several data sets with duplicate points and ...

A Robust Algorithm for Characterizing Anisotropic Local ...
755 College Road East, Princeton, NJ 08540, USA. 2. INRIA Rhône- ...... lung walls and near rib structures. All the 14 ... Academic Press, San Diego, 1990. 5.

An improved memetic algorithm using ring neighborhood ... - GitHub
4, 5, 6, 7 that the con- vergence to the known optimal result of test functions is very fast and most test functions have been converged after around. 1 × 105 FEs.

An improved memetic algorithm using ring neighborhood ... - GitHub
Nov 29, 2013 - The main motivation of using ring neighborhood topology is to provide a good ... mine the choice of solutions for local refinements, by utiliz- ...... 93,403 g08. 2,755. 2,990 g09. 13,455. 23,990 g10. 95,788. 182,112 g11. 1,862.

W-AlignACE: an improved Gibbs sampling algorithm based on more ...
Computer Science and Technology, Tsinghua University, Beijing, China and 3Department of Computer Science and Engineering ...... Singapore Ministry of Education and T.J.'s research is sup- ported by ... Genet, 27, 167–171. Cherry,J. et al.

W-AlignACE: an improved Gibbs sampling algorithm based on more ...
learning an accurate PWM to characterize the binding sites of a specific TF ... W-AlignACE, is compared with three other programs (AlignACE,. MDscan and ..... relative entropy (i.e. Kullback–Leibler distance) of binding sites with respect to the ..

An Improved FloatBoost Algorithm for Naïve Bayes Text ...
20-Newsgroups consists of 20,000 Usenet articles collected by K. Lang from 20 different newsgroups. For this data set, about 70% documents in each newsgroup are used for training (700 documents per class), while left documents are used for testing. (

An engineered anisotropic nanofilm with ...
S3) at ∼85 Hz. Drop motion was recorded with a Samsung HMX-H106 digital camcorder. Under random ... for the advancing and receding contact angles into a Mathematica notebook, included with the ... New Phytol. 184, 988–1002 (2009).

An Evolutionary Algorithm for Homogeneous ...
fitness and the similarity between heterogeneous formed groups that is called .... the second way that is named as heterogeneous, students with different ...

An Algorithm for Implicit Interpolation
More precisely, we consider the following implicit interpolation problem: Problem 1 ... mined by the sequence F1,...,Fn and such that the degree of the interpolants is at most n(d − 1), ...... Progress in Theoretical Computer Science. Birkhäuser .

An Adaptive Fusion Algorithm for Spam Detection
An email spam is defined as an unsolicited ... to filter harmful information, for example, false information in email .... with the champion solutions of the cor-.

An Algorithm for Implicit Interpolation
most n(d − 1), where d is an upper bound for the degrees of F1,...,Fn. Thus, al- though our space is ... number of arithmetic operations required to evaluate F1,...,Fn and F, and δ is the number of ...... Progress in Theoretical Computer Science.

An Adaptive Fusion Algorithm for Spam Detection
adaptive fusion algorithm for spam detection offers a general content- based approach. The method can be applied to non-email spam detection tasks with little ..... Table 2. The (1-AUC) percent scores of our adaptive fusion algorithm AFSD and other f

An Algorithm for Nudity Detection
importance of skin detection in computer vision several studies have been made on the behavior of skin chromaticity at different color spaces. Many studies such as those by Yang and Waibel (1996) and Graf et al. (1996) indicate that skin tones differ

An Improved Control Law Using HVDC Systems for ...
Aug 28, 2013 - Systems for Frequency Control. Jing Dai1. Gilney Damm2 ... information. ... This lead to a smaller degree of primary reserve sharing, than the ...

An Improved Degree Based Condition for Hamiltonian ...
Lenin Mehedy1, Md. Kamrul Hasan1 and Mohammad Kaykobad2. 1Department of Computer Engineering, Kyung Hee University, South Korea. 2Department of Computer Science and Engineering, North South University, Dhaka, Bangladesh. Email: 1{lenin, kamrul}@oslab

An Improved Likelihood Model for Eye Tracking
Dec 6, 2005 - This property makes eye tracking systems a unique and effective tool for disabled people ...... In SPIE Defense and Security Symposium,. Automatic ... Real-time eye, gaze, and face pose tracking for monitoring driver vigilance.

An Improved Particle Swarm Optimization for Prediction Model of ...
An Improved Particle Swarm Optimization for Prediction Model of. Macromolecular Structure. Fuli RONG, Yang YI,Yang HU. Information Science School ...

AN IMPROVED CONSENSUS-LIKE METHOD FOR ...
{haihua xu,zhujie}@sjtu.edu.cn, [email protected], [email protected]. ABSTRACT ... for lattice rescoring and lattice-based system combination, versus baselines such .... similar approximations as used in Minimum Phone/Word Error.

OMEGA: An Improved Gasoline Blending System for ... - EBSCOhost
refinery data bases and on-line data acquisition and exploits detailed nonlinear models of gasoline attributes. Texaco uses. OMEGA in all seven US refineries ...