Difference between revisions of "Projects:SphericalDemons"

From NAMIC Wiki
Jump to: navigation, search
(New page: Back to NA-MIC Collaborations, MIT Algorithms We propose a unified framework for computing atlases from manual...)
 
 
(52 intermediate revisions by 4 users not shown)
Line 1: Line 1:
−
Back to [[NA-MIC_Internal_Collaborations:StructuralImageAnalysis|NA-MIC Collaborations]], [[Algorithm:MIT|MIT Algorithms]]
+
Back to [[NA-MIC_Internal_Collaborations:StructuralImageAnalysis|NA-MIC Collaborations]], [[Algorithm:MIT|MIT Algorithms]]
 +
__NOTOC__
 +
= Spherical Demons =
  
−
We propose a unified framework for computing atlases from manually labeled data sets at various degrees of “sharpness” and the joint registration and segmentation of a new brain with these atlases. In non-rigid registration, the tradeoff between warp regularization and image fidelity is typically set empirically. In segmentation, this leads to a probabilistic atlas of arbitrary “sharpness”: weak regularization results in well-aligned training images, producing a “sharp” atlas; strong regularization yields a “blurry” atlas. We study the effects of this tradeoff in the context of cortical surface parcellation, but the framework applies to volume registration as well. This is an important question because of the increasingly availability of atlases in public databases and the development of registration algorithms separate from the atlas construction process.
+
We present the fast Spherical Demons algorithm for registering two spherical images. By exploiting spherical vector spline
 +
interpolation theory, we show that a large class of regularizers for the modified demons objective function can be efficiently
 +
approximated on the sphere using convolution. Based on the one parameter subgroups of diffeomorphisms, the resulting registration is
 +
diffeomorphic and fast -- registration of two cortical mesh models with more than 100k nodes takes less than 5 minutes, comparable to the fastest surface registration algorithms. Moreover, the accuracy of our method compares favorably to the popular FreeSurfer
 +
registration algorithm. We validate the technique in two different settings: (1) parcellation in a set of in-vivo cortical surfaces and (2) Brodmann area localization in ex-vivo cortical surfaces.
  
 
= Description =
 
= Description =
−
In image registration, one usually optimizes the objective function with two parts. The first term is the similarity between images. The second term regularizes the warp. The smoothness parameter that weights the second term determines the tradeoff between the similarity measure and the regularization. In Atlas-based segmentation, one is given a set of labeled training images. The training images are co-registered to a common space. An atlas that summarizes the information between the image features and the labels is computed in this common space. This atlas is used to segment and normalize a new image.
+
Motivated by the spherical representation of the cerebral cortex, this work deals with the problem of registering spherical images. Cortical folding patterns are correlated with both cytoarchitectural and functional regions. In group studies of cortical structure and function, determining corresponding folds across subjects is therefore important.
  
−
[[Image:RegSeg.png|center|400px|]]
+
Unfortunately, many spherical warping algorithms are computationally expensive. One reason is the need for invertible deformations that preserve the topology of structural or functional regions across subjects. In this work, we take the approach, previously demonstrated in the Euclidean space [3], of restricting the deformation space to be a composition of diffeomorphisms, each of which is parameterized by a stationary velocity field. In each iteration, the algorithm greedily seeks the best diffeomorphism to be composed with the current transformation, resulting in much faster updates.  
  
−
We employ a generative model for the joint registration and segmentation of images (see figures below). The atlas construction process arises naturally as estimation of the model parameters. This framework allows the computation of unbiased atlases from manually labeled data at various degrees of "sharpness", as well as the joint registration and segmentation of a novel brain in a consistent manner.
+
Another challenge in registration is the tradeoff between the image similarity measure and the regularization in the objective function. Since most regularizations favor smooth deformations, the gradient computation is complicated by the need to take into account the deformation in neighboring regions. For Euclidean images, the demons objective function facilitates a fast two-step optimization where the second step handles the warp regularization via a single convolution with a smoothing filter [2,3]. Based on spherical vector spline interpolation theory and other differential geometric tools, we show that the two-stage optimization procedure of the demons algorithm can be efficiently applied on the sphere.
  
−
<center>
+
[[Image:CoordinateChart.png | center | 400px]]
−
<table>
 
−
<tr>
 
−
<td>
 
−
[[Image:JointRegSeg.png|thumb|center|420px| Joint Registration-Segmentation]]
 
−
</td>
 
−
<td>
 
−
[[Image:GraphicalModel.png|thumb|center|400px|"A" is an atlas used to generate the label map L' in some universal atlas space. The atlas A and label map L' generate image I'. S is the smoothness parameter that generates random warp field R. This warp is then applied to the label map L' and image I' to create the label map L and the image I. We assume the label map L is available for the training images, but not for the test image. The image I is observed in both training and test cases.]]
 
−
</td>
 
−
</tr>
 
−
</table>
 
−
</center>
 
  
−
We use the generative model to compute atlases from manually labeled data at various degrees of “sharpness” and the joint registration-segmentation of a new brain with these atlases. Using this framework, we investigate the tradeoff between warp regularization and image fidelity, i.e. the smoothness of the new subject warp and the sharpness of the atlas. We compare three special cases of our framework, namely:
+
== Experimental Results ==
 +
We use two sets of experiments to compare the accuracy of Spherical Demons and FreeSurfer [1]. The FreeSurfer registration algorithm uses the same similarity measure as Spherical Demons, but penalizes for metric and areal distortion. Its runtime is more than an hour while our runtime is less than 5 minutes.
  
−
(1) Progressive registration of a new brain to increasingly “sharp” atlases using increasingly flexible warps, by initializing each registration stage with the optimal warps from a “blurrier” atlas. We call this multiple atlases, multiple warp scales (MAMS).
+
== (1) Parcellation of In-Vivo Cortical Surfaces ==
  
−
(2) Progressive registration to a single atlas with increasingly flexible warps. We call this single atlas, multiple warp scales (SAMS).
+
We consider a set of 39 left and right cortical surface models extracted from in-vivo MRI. Each surface is spherically parameterized and represented as a spherical image with geometric features at each vertex (e.g., sulcal depth and curvature). Both hemispheres are manually parcellated by a neuroanatomist into 35 major sulci and gyri. We validate our algorithm in the context of automatic cortical parcellation.
  
−
(3) Registration to a single atlas with fixed constrained warps. We call this single atlas, single warp scale (SASS).
+
We co-register all 39 spherical images of cortical geometry with Spherical Demons by iteratively building an atlas and registering
 +
the surfaces to the atlas. The atlas consists of the mean and variance of cortical geometry. We then perform cross-validation parcellation 4 times, by leaving out subjects 1 to 10, training a classifier using the remaining subjects, and using it to classify subjects 1 to 10. We repeat with subjects 11-20, 21-30 and 31-39. We also perform registration and cross-validation with the FreeSurfer algorithm [1] using the same features and parcellation algorithm. Once again, the atlas consists of the mean and variance of cortical geometry.
  
−
= Experimental Results =
 
  
−
We use dice as the measure of segmentation quality. From the graph below, we note that the optimal algorithms correspond to a unique balance between atlas “sharpness” and warp regularization. Our experiments show that the optimal parameter values that correspond to this balance can be determined using cross-validation. The optimal parameter values are robust across subjects, and the same for both co-registration of the training data and registration of a new subject. This suggests that a single atlas at an optimal sharpness is sufficient to achieve the best segmentation results. Furthermore, our experiments also suggest that segmentation accuracy is tolerant up to a small mismatch between atlas sharpness and warp smoothness.
+
The average Dice measure (defined as the ratio of cortical surface area with correct labels to the total surface area averaged over the test set) on the left hemisphere is 88.9 for FreeSurfer and 89.6 for Spherical Demons. While the improvement is not big, the
−
 
+
difference is statistically significant for a one-sided t-test with the Dice measure of each subject treated as an independent sample (p = 2e-6). On the right hemisphere, FreeSurfer obtains a Dice of 88.8 and Spherical Demons achieves 89.1. Here, the improvement is smaller, but still statistically significant (p = 0.01).
−
[[Image:AvgResults.jpg|thumb|center|300px|Plot of Dice as a function of the warp smoothness S. Note that S is on a log scale. <math>\alpha</math> corresponds to the sharpness of the atlas used.]]
 
−
 
 
−
In the figure below, we display the percentage improvement of SASS over FreeSurfer [1,2]. For each of the 35 structures for each hemisphere, we perform a one-sided paired-sampled t-test between SASS and FreeSurfer, where each subject is considered a sample. We use the False Discovery Rate (FDR) to correct for multiple comparisons. In the left hemisphere, SASS achieves statistically significant improvement over FreeSurfer for 17 structures (FDR < 0.05), while the remaining structures yield no statistical difference. In the right hemisphere, SASS achieves improvement for 11 structures (FDR < 0.05), while the remaining structures yield no statistical difference. The p-values for the left and right hemispheres are pooled together for the False Discovery Rate analysis.
 
  
 
<center>
 
<center>
Line 43: Line 36:
 
<tr>
 
<tr>
 
<td>
 
<td>
−
[[Image:Lh.PercentImprove1.png|thumb|center|150px|Left Medial]]
+
[[Image:SD.Lh.PercentImprove.lat.png|thumb|center|150px|Left Medial]]
 
</td>
 
</td>
 
<td>
 
<td>
−
[[Image:Lh.PercentImprove2.png|thumb|center|150px|Left Lateral]]
+
[[Image:SD.Lh.PercentImprove.med.png|thumb|center|150px|Left Lateral]]
 
</td>
 
</td>
 
<td>
 
<td>
−
[[Image:Rh.PercentImprove2.png|thumb|center|150px|Right Lateral]]
+
[[Image:SD.Rh.PercentImprove.lat.png|thumb|center|150px|Right Lateral]]
 
</td>
 
</td>
 
<td>
 
<td>
−
[[Image:Rh.PercentImprove1.png|thumb|center|142px|Right Medial]]
+
[[Image:SD.Rh.PercentImprove.med.png|thumb|center|142px|Right Medial]]
 
</td>
 
</td>
 
</tr>
 
</tr>
Line 58: Line 51:
 
</center>
 
</center>
  
 +
Because the average Dice can be deceiving by suppressing small structures, we analyze the segmentation accuracy per structure. On
 +
the left (right) hemisphere, the segmentations of 16 (8) structures are statistically significantly improved by Spherical Demons with respect to FreeSurfer, while no structure got worse (FDR = 0.05). The above figure shows the percentage improvement
 +
of individual structures. Parcellation results suggest that our registration is at least as accurate as FreeSurfer.
  
−
[1] Fischl, Sereno and Dale. High-resolution intersubject averaging and a coordinate system for the cortical surface. Human Brain Mapping, 8(4):272--284, 1999
+
== (2) Brodmann Areas Localization on Ex-vivo Cortical Surfaces ==
 +
 
 +
In this experiment, we evaluate the registration accuracy on ten human brains analyzed histologically postmortem. The histological sections were aligned to postmortem MR with nonlinear warps to build a 3D volume. Eight manually labeled Brodmann areas from histology were sampled onto each hemispheric surface model and sampling errors were manually corrected. Brodmann areas are cyto-architectonically defined regions closely related to cortical function.
 +
 
 +
It has been shown that nonlinear surface registration of cortical folds can significantly improve Brodmann area overlap across
 +
different subjects. Registering the ex-vivo surfaces is more difficult than in-vivo surfaces because the reconstructed volumes are extremely noisy, resulting in noisy geometric features.
 +
 
 +
We co-register the ten surfaces to each other by iteratively building an atlas and registering the surfaces to the atlas. We
 +
compute the average distance between the boundaries of the Brodmann areas for each pair of registered subjects. We perform a permutation test to test for statistical significance. Spherical Demons improves the alignment of 5 (2) Brodmann areas on the left (right) hemisphere (FDR = 0.05) compared with FreeSurfer and no structure gets worse. These results suggest that the Spherical Demons algorithm is at least as accurate as FreeSurfer in aligning Brodmann areas.
 +
 
 +
= Code =
 +
 
 +
Matlab code is currently available at http://yeoyeo02.googlepages.com/sphericaldemonsrelease
 +
 
 +
We are currently working in collaboration with Kitware to make the code available in ITK.
  
−
[2] Fischl, van der Kouwe, Destrieux, Halgren, Segonne, Salat, Busa, Seidman, Goldstein, Kennedy, Caviness, Makris, Rosen and Dale. Automatically Parcellating the Human cerebral Cortex. Cerebral Cortex, 14:11--22, 2004
 
  
−
= Key Investigators =
 
  
−
* MIT Algorithms: [[http://people.csail.mit.edu/ythomas/ | B.T. Thomas Yeo]], Mert Sabuncu, Rahul Desikan, Bruce Fischl, Polina Golland
+
[1] B. Fischl, M. Sereno, R. Tootell, and A. Dale. High-resolution Intersubject Averaging and a Coordinate System for the
 +
Cortical Surface. Human Brain Mapping, 8(4):272–284, 1999.
  
−
= Publications =
+
[2] J. Thirion. Image Matching as a Diffusion Process: an Analogy with Maxwell’s Demons. Medical Image Analysis,
−
''In Print''
+
2(3):243–260, 1998.
  
−
* [http://www.na-mic.org/pages/Special:Publications?text=Projects%3ARegistrationRegularization&submit=Search&words=all&title=checked&keywords=checked&authors=checked&abstract=checked&sponsors=checked&searchbytag=checked| NA-MIC Publications Database]
+
[3] T. Vercauteren, X. Pennec, A. Perchant, and N. Ayache. Non-parametric Diffeomorphic Image Registration with the
 +
Demons Registration. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted
 +
Intervention (MICCAI), volume 4792 of LNCS, pages 319–326, 2007.
  
−
''In Press''
+
= Key Investigators =
  
−
* B.T.T. Yeo, M.R. Sabuncu, R. Desikan, B. Fischl, P. Golland. "Effects of Registration Regularization and Atlas Sharpness on Segmentation Accuracy". Accepted to Medical Image Analysis, 2008.
+
* MIT: [[http://people.csail.mit.edu/ythomas/ | B.T. Thomas Yeo]], Mert Sabuncu, Polina Golland.
 +
* Harvard: Bruce Fischl.  
 +
* INRIA: Nicholas Ayache.  
 +
* Mauna Kea Technologies: Tom Vercauteren.
 +
* Kitware: Luis Ibanez, Michel Audette.
 +
 
 +
= Publications =
  
 +
* [http://www.na-mic.org/publications/pages/display?search=Projects:SphericalDemons&submit=Search&words=all&title=checked&keywords=checked&authors=checked&abstract=checked&sponsors=checked&searchbytag=checked| NA-MIC Publications Database on Spherical Demons: Fast Surface Registration]
  
  
 
[[Category: Registration]] [[Category:Segmentation]]
 
[[Category: Registration]] [[Category:Segmentation]]

Latest revision as of 20:04, 11 May 2010

Home < Projects:SphericalDemons
Back to NA-MIC Collaborations, MIT Algorithms

Spherical Demons

We present the fast Spherical Demons algorithm for registering two spherical images. By exploiting spherical vector spline interpolation theory, we show that a large class of regularizers for the modified demons objective function can be efficiently approximated on the sphere using convolution. Based on the one parameter subgroups of diffeomorphisms, the resulting registration is diffeomorphic and fast -- registration of two cortical mesh models with more than 100k nodes takes less than 5 minutes, comparable to the fastest surface registration algorithms. Moreover, the accuracy of our method compares favorably to the popular FreeSurfer registration algorithm. We validate the technique in two different settings: (1) parcellation in a set of in-vivo cortical surfaces and (2) Brodmann area localization in ex-vivo cortical surfaces.

Description

Motivated by the spherical representation of the cerebral cortex, this work deals with the problem of registering spherical images. Cortical folding patterns are correlated with both cytoarchitectural and functional regions. In group studies of cortical structure and function, determining corresponding folds across subjects is therefore important.

Unfortunately, many spherical warping algorithms are computationally expensive. One reason is the need for invertible deformations that preserve the topology of structural or functional regions across subjects. In this work, we take the approach, previously demonstrated in the Euclidean space [3], of restricting the deformation space to be a composition of diffeomorphisms, each of which is parameterized by a stationary velocity field. In each iteration, the algorithm greedily seeks the best diffeomorphism to be composed with the current transformation, resulting in much faster updates.

Another challenge in registration is the tradeoff between the image similarity measure and the regularization in the objective function. Since most regularizations favor smooth deformations, the gradient computation is complicated by the need to take into account the deformation in neighboring regions. For Euclidean images, the demons objective function facilitates a fast two-step optimization where the second step handles the warp regularization via a single convolution with a smoothing filter [2,3]. Based on spherical vector spline interpolation theory and other differential geometric tools, we show that the two-stage optimization procedure of the demons algorithm can be efficiently applied on the sphere.

CoordinateChart.png

Experimental Results

We use two sets of experiments to compare the accuracy of Spherical Demons and FreeSurfer [1]. The FreeSurfer registration algorithm uses the same similarity measure as Spherical Demons, but penalizes for metric and areal distortion. Its runtime is more than an hour while our runtime is less than 5 minutes.

(1) Parcellation of In-Vivo Cortical Surfaces

We consider a set of 39 left and right cortical surface models extracted from in-vivo MRI. Each surface is spherically parameterized and represented as a spherical image with geometric features at each vertex (e.g., sulcal depth and curvature). Both hemispheres are manually parcellated by a neuroanatomist into 35 major sulci and gyri. We validate our algorithm in the context of automatic cortical parcellation.

We co-register all 39 spherical images of cortical geometry with Spherical Demons by iteratively building an atlas and registering the surfaces to the atlas. The atlas consists of the mean and variance of cortical geometry. We then perform cross-validation parcellation 4 times, by leaving out subjects 1 to 10, training a classifier using the remaining subjects, and using it to classify subjects 1 to 10. We repeat with subjects 11-20, 21-30 and 31-39. We also perform registration and cross-validation with the FreeSurfer algorithm [1] using the same features and parcellation algorithm. Once again, the atlas consists of the mean and variance of cortical geometry.


The average Dice measure (defined as the ratio of cortical surface area with correct labels to the total surface area averaged over the test set) on the left hemisphere is 88.9 for FreeSurfer and 89.6 for Spherical Demons. While the improvement is not big, the difference is statistically significant for a one-sided t-test with the Dice measure of each subject treated as an independent sample (p = 2e-6). On the right hemisphere, FreeSurfer obtains a Dice of 88.8 and Spherical Demons achieves 89.1. Here, the improvement is smaller, but still statistically significant (p = 0.01).

Left Medial
Left Lateral
Right Lateral
Right Medial

Because the average Dice can be deceiving by suppressing small structures, we analyze the segmentation accuracy per structure. On the left (right) hemisphere, the segmentations of 16 (8) structures are statistically significantly improved by Spherical Demons with respect to FreeSurfer, while no structure got worse (FDR = 0.05). The above figure shows the percentage improvement of individual structures. Parcellation results suggest that our registration is at least as accurate as FreeSurfer.

(2) Brodmann Areas Localization on Ex-vivo Cortical Surfaces

In this experiment, we evaluate the registration accuracy on ten human brains analyzed histologically postmortem. The histological sections were aligned to postmortem MR with nonlinear warps to build a 3D volume. Eight manually labeled Brodmann areas from histology were sampled onto each hemispheric surface model and sampling errors were manually corrected. Brodmann areas are cyto-architectonically defined regions closely related to cortical function.

It has been shown that nonlinear surface registration of cortical folds can significantly improve Brodmann area overlap across different subjects. Registering the ex-vivo surfaces is more difficult than in-vivo surfaces because the reconstructed volumes are extremely noisy, resulting in noisy geometric features.

We co-register the ten surfaces to each other by iteratively building an atlas and registering the surfaces to the atlas. We compute the average distance between the boundaries of the Brodmann areas for each pair of registered subjects. We perform a permutation test to test for statistical significance. Spherical Demons improves the alignment of 5 (2) Brodmann areas on the left (right) hemisphere (FDR = 0.05) compared with FreeSurfer and no structure gets worse. These results suggest that the Spherical Demons algorithm is at least as accurate as FreeSurfer in aligning Brodmann areas.

Code

Matlab code is currently available at http://yeoyeo02.googlepages.com/sphericaldemonsrelease

We are currently working in collaboration with Kitware to make the code available in ITK.


[1] B. Fischl, M. Sereno, R. Tootell, and A. Dale. High-resolution Intersubject Averaging and a Coordinate System for the Cortical Surface. Human Brain Mapping, 8(4):272–284, 1999.

[2] J. Thirion. Image Matching as a Diffusion Process: an Analogy with Maxwell’s Demons. Medical Image Analysis, 2(3):243–260, 1998.

[3] T. Vercauteren, X. Pennec, A. Perchant, and N. Ayache. Non-parametric Diffeomorphic Image Registration with the Demons Registration. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), volume 4792 of LNCS, pages 319–326, 2007.

Key Investigators

  • MIT: [| B.T. Thomas Yeo], Mert Sabuncu, Polina Golland.
  • Harvard: Bruce Fischl.
  • INRIA: Nicholas Ayache.
  • Mauna Kea Technologies: Tom Vercauteren.
  • Kitware: Luis Ibanez, Michel Audette.

Publications

Retrieved from "https://www.na-mic.org/w/index.php?title=Projects:SphericalDemons&oldid=52214"