# Additional Results for Joint Segmentation/Registration model by shape alignement via Weighted Total Variation Minimization and Nonlinear Elasticity

## Abstract

This method falls within the scope of joint segmentation-registration using nonlinear elasticity principles. Saint Venant-Kirchhoff materials being the simplest hyperelastic materials (hyperelasticity being a suitable framework when dealing with large and nonlinear deformations), we propose viewing the shapes to be matched as such materials. Then we introduce a variational model combining a measure of dissimilarity based on weighted total variation and a regularizer based on the stored energy function of a Saint Venant-Kirchhoff material. Adding a weighted total variation based criterion enables to align the edges of the objects even if the modalities are different.

We recall the functionnal to minimize : \begin{align} \bar{I}_{\gamma}(\varphi,V,\tilde{T})=&{\mbox{var}}_g\,\tilde{T}+\dfrac{\nu}{2}\,\|T(\varphi)-R\|_{L^2(\Omega)}^2+\displaystyle{\int_{\Omega}}\,QW(V)\,dx\nonumber \\ &+\dfrac{\gamma}{2}\,\|V-\nabla \varphi\|_{L^2(\Omega,M_2)}^2+\gamma\,\|\tilde{T}-T \circ \varphi\|_{L^1(\Omega)}. \label{le-guyader-equation8} \end{align} with $$QW(\xi)=\left\{\begin{array}{ll} W(\xi)\,\,\,{\mbox{if \|\xi\|^2 \geq 2\dfrac{\lambda+\mu}{\lambda+2\mu}}},\\ \Psi({\mbox{det}}\,\xi)\,\,\,{\mbox{if \|\xi\|^2 < 2\dfrac{\lambda+\mu}{\lambda+2\mu}}}, \end{array}\right.$$ We set $$\widehat{\mathcal{W}}={\mbox{Id}}+W_0^{1,2}(\Omega,\mathbb{R}^2)$$ and $$\widehat{\mathcal{\chi}}=\left\{V \in L^{4}(\Omega,M_2(\mathbb{R}))\right\}$$. The decoupled problem consists in minimizing $$\bar{I}_{\gamma}$$ on $$\widehat{\mathcal{W}} \times \widehat{\mathcal{\chi}}\times BV(\Omega,g)$$

## Results on synthetic images

### Letter_C

First, the method is applied on an academic example taken from [1] for mapping a disk to the letter C, demonstrating the ability of the algorithm to handle large deformations. Note that with linear elasticity model, diffusion model or curvature-based model, registration cannot be successfully accomplished (see [2]). As in [1], the right part of the disk is stretched into the shape of the interior edge of the letter C, and then moves outward to align the interior boundary of the letter C. Nevertheless, our deformation field is smoother (see in particular [1, p. 88]).

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse Deformation grid

Displacement vector field

min det $$\nabla \varphi$$=0.002, max det $$\nabla \varphi$$=2.32.
$$\Delta MI = 12\%$$.

### Triangle

Another toy example is provided to emphasize again the capability of the model to generate large deformations even on data corrupted by noise. The algorithm produces both a smooth deformation field and a simplified (thus here denoised) version of the Reference image allowing for its segmentation.

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse Deformation grid

Displacement vector field

min det $$\nabla \varphi$$=0.37, max det $$\nabla \varphi$$=2.25.
$$\Delta MI = 15\%$$.

## Results on Mouse Atlas

Then the method is applied on medical images with the goal to map a 2D slice of mouse brain gene expression data (Template T) to its corresponding 2D slice of the mouse brain atlas, in order to facilitate the integration of anatomic, genetic and physiologic observations from multiple subjects in a common space. Since genetic mutations and knock-out strains of mice provide critical models for a variety of human diseases, such linkage between genetic information and anatomical structure is important. The data are provided by the Center for Computational Biology, UCLA. The mouse atlas acquired from the LONI database was pre-segmented. The gene expression data were segmented manually to facilitate data processing in other applications. Some algorithms have been developed to automatically segment the brain area of gene expression data. The non-brain regions have been removed to produce better matching. Our method qualitatively performs as the one in [3] and produces a smooth deformation field but also provides both a simplified version of the Reference image and its segmentation.

### Atlas11

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

$$\tilde{T}$$

min det $$\nabla \varphi$$=0.63, max det $$\nabla \varphi$$=1.28.
$$\Delta MI = 6\%$$.

### Atlas12

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

Ttilde

min det $$\nabla \varphi$$=0.68, max det $$\nabla \varphi$$=1.42.
$$\Delta MI = 14\%$$.

### Atlas06

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

Ttilde

min det $$\nabla \varphi$$=0.46, max det $$\nabla \varphi$$=1.58.
$$\Delta MI = 35\%$$.

### Atlas07

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

Ttilde

min det $$\nabla \varphi$$=0.17, max det $$\nabla \varphi$$=1.45.
$$\Delta MI = 24\%$$.

### Atlas08

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

Ttilde

min det $$\nabla \varphi$$=0.052, max det $$\nabla \varphi$$=2.31.
$$\Delta MI = 41\%$$.

## Brain

The method has also been applied to complex slices of brain data (courtesy of Laboratory Of Neuro-Imaging, UCLA). We aim to register a torus to the slice of brain with topology preservation to demonstrate the ability of the algorithm to handle complex topologies. The results are very satisfactory on this example since the deformed Template matches very well the convolutions of the brain. The interior contour of the right part of the torus moves to the upper boundary of the hole, while the exterior contour moves towards the upper envelope, entailing large deformations since the thickness of this part of the brain is much greater than the thickness of the torus.

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

$$\tilde{T}$$

min det $$\nabla \varphi$$=0.006, max det $$\nabla \varphi$$=14.8.
$$\Delta MI = 50\%$$.

## MRI images of cardiac cycle

Finally, numerical simulations on MRI images of a patient cardiac cycle have been carried out. We were supplied with a whole cardiac MRI examination of a patient (courtesy of the LITIS, University of Rouen, France). It is made of 280 images divided into 14 levels of slice and 20 images per cardiac cycle. The numbering of the images goes from 0 to 279, and includes both the slice number and the time index. The image 0 is set at the upper part of the heart and the sequence from image 0 to image 19 contains the whole cardiac cycle for this slice. The sequence from images 20 to 39 contains the whole cardiac cycle for the slice underneath the previous one and so on. A cardiac cycle is composed of a contraction phase (40% of the cycle duration), followed by a dilation phase (60% of the cycle duration). The first image of the sequence (frames 0, 20, 40, etc.) is when the heart is most dilated (end diastole - ED) and the 8 th of the sequence (end systole - ES) is when the heart is most contracted. It thus seemed relevant, in order to assess the accuracy of the proposed algorithm in handling large deformations, to register the pair ED-ES. Besides, due to the patient’s breathing, images from a slice to another are not stackable (whereas they should be) so we also registered pairs of the form 120-140.

### Heart ED-ES (images 80 to 88)

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

$$\tilde{T}$$

min det $$\nabla \varphi$$=0.01, max det $$\nabla \varphi$$=3.83.
$$\Delta MI = 33\%$$.

### Heart ED-ES (images 100 to 108)

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

$$\tilde{T}$$

min det $$\nabla \varphi$$=0.005, max det $$\nabla \varphi$$=2.31.
$$\Delta MI =27\%$$.

### Heart 120-140

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

$$\tilde{T}$$

min det $$\nabla \varphi$$=0.18, max det $$\nabla \varphi$$=2.92.
$$\Delta MI = 27\%$$.

### Heart 160-180

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

$$\tilde{T}$$

min det $$\nabla \varphi$$=0.11, max det $$\nabla \varphi$$=3.06.
$$\Delta MI = 24\%$$.

### Heart 80-81

We also have tried to register two consecutive images, however, the deformation being very small, it is quite difficult to appreciate the result.

Template

Reference

Deformed Template

Segmented Reference

Deformation grid

Inverse deformation

Displacement vector field

$$\tilde{T}$$

min det $$\nabla \varphi$$=0.65, max det $$\nabla \varphi$$=1.89.
$$\Delta MI = 12\%$$.

## Comparison with prior related works

In [3], Lin et al. first review the most common and simplest regularization terms (diffusion, biharmonic, linear elasticity models) that lead to linear terms with respect to derivatives in the Euler-Lagrange equations. One of their conclusions is that the biharmonic model is more comparable to the nonlinear elasticity model, which motivates us to further examine its behaviour compared with our model. First, we have compared the obtained results on the torus-brain example by replacing the nonlinear-elasticity-based regularizer by the biharmonic one and by removing the weighted total variation in order to assess its relevance. The obtained deformed Template exhibits artefacts on the boundary of the brain slice, and the hole is not satisfactorily reproduced after 40000 iterations for this algorithm versus 100 iterations for our model.

Deformed Template

The result is better adding the weighted total variation in particular on the boundary, but the hole is not reproduced either.

Deformed Template

## References

1. G. E. Christensen, Deformable shape models for anatomy, PhD thesis, Washing- ton University, Sever Institute of technology, USA, 1994.
2. J. Modersitzki, Numerical Methods for Image Registration, Oxford University Press, 2004
3. T. Lin, C. Le Guyader, I. Dinov, P. Thompson, A. Toga, and L. Vese, Gene Expression Data to Mouse Atlas Registration Using a Nonlinear Elasticity Smoother and Landmark Points Constraints, Washington, J. Sci. Comput. 50 (2012), pp. 586– 609.