Weighted Metamorphosis
For registration of images with different topologies.
Anton François
supervised by
Joan Glaunès & Pietro Gori
The Registration Problem
The Registration Problem
- Necrosis : topological difference
- Mass effect : morphological difference
- Oedema : appearance difference
- Brain natural differences
MNI template
The Registration Problem
Metamorphosis & LDDMM
We aim at finding \( (v_t)_{t\in [0,1]}\) and \( (z_t)_{t \in [0,1]}\) which minimise
$$ E(I,v) = \frac 12 \| I_1 - T \|_{L^2}^2 + Reg(\bullet)$$
$$\quad s.t.: \dot I_t = -\langle \nabla I_t, v_t \rangle + \mu z_t, \quad I_0 = S, \mu \in \mathbb R^+$$
Metamorphosis : LDDMM + Intensity changes
LDDMM : Register objects with the help of diffeomorphic deformations inducing a metric between images.
Metamorphosis & LDDMM
$$\left\{\begin{array}{rl}v_t &= -\frac{\rho}{\mu} K_\sigma \star (z_t \nabla I_t)\\ \dot z_t &= -\quad \nabla \cdot (z_t v_t) \\ \dot I_t &= -\langle v_t , \nabla I_t\rangle + \mu z_t\end{array}\right.$$
Advection equation
Continuity equation
To find the minimum of this cost :
$$ E_\mathrm M (I,v) = \frac12 \left\| I_1 - T \right\|_2^2 + \frac12\int_0^1 \|v_t\|_V^2 + \rho \|z_t\|_{L_2}^2 dt $$
We do geodesic shooting, integrating over this set of geodesic equations :
Source
Target
LDDMM
Metamorphsosis
Weighted Metamorphosis
$$\left\{\begin{array}{rl}v_t &= -\frac{\rho}{\mu} K_\sigma \star (z_t \nabla I_t)\\ \dot z_t &= -\quad \nabla \cdot (z_t v_t) \\ \dot I_t &= -\langle v_t , \nabla I_t\rangle + \mu M_t z_t\end{array}\right.$$
Let \((M_t)_{t\in [0,1]}\) be a continuous temporal mask.
To find the minimum of this cost :
$$ E_{\mathrm{WM}}(I,v) = \frac12 \left\| I_1 - T \right\|_2^2 + \frac12\int_0^1 \|v_t\|_V^2 + \rho \langle z_t, M_t z_t \rangle_{L^2} dt $$
We have to integrate over the set of geodesic equations :
Source
Target
LDDMM
Metamorphsosis
WM
Weighted Metamorphosis
with a growing mask.
1. We set \(M_1\) as the topological difference segmentation.
2. We initialise \(M_0\) as a small ball at its center.
3. We register \(M_0\) to \(M_1\) using LDDMM.
Source
Target
WM
WM w. Growing mask
Real data :
Results in 3D
Target
Target
github.com/antonfrancois/Demeter_metamorphosis
Good for registering object with a growing topology change.
A bit slower than DL methods at inference time.
Works on individual images.
WBIR-WM
By Anton FRANCOIS
WBIR-WM
- 59