next up previous
Next: Summary Up: Multi-Variate Gaussian Intensity Prior Previous: Marginalisation over Areas of

Marginalisation over Partial Volume Parameters

Pure partial volume parameters can again be split into those associated with shapes of interest, $ \alpha_{pv,i}$, and those associated with areas of no interest, $ \alpha_{pv,u}$. The former have non-flat priors and the latter have flat priors. Integration of the former requires more care, as they interact with the multi-variate prior. In order to do this, rewrite the appropriate matrices as

$\displaystyle \alpha = \left[ \begin{array}{c} \alpha_{pv,i} \\ \alpha_{in} \\ ...
...^{\mathrm{\textsf{T}}}\\ Q_{b} & & Q_{cross} & & Q_{null} \end{array} \right].
$

Therefore the posterior integrations take the form

    $\displaystyle \int \exp \left( \frac{-\beta}{2} (Y - G \alpha)^{\mathrm{\textsf...
...Y - G \alpha) - \lambda \alpha^{\mathrm{\textsf{T}}}Q \alpha \right) \, d\alpha$  
    $\displaystyle \qquad \qquad = \int \exp \left( \frac{-\beta}{2} (Y - G_{pv,i} \...
...ambda \alpha_{pv,i}^{\mathrm{\textsf{T}}}Q_{pv} \alpha_{pv,i} \right) \; \times$  
    $\displaystyle \qquad \qquad \qquad \exp \left( \frac{-\beta}{2} (\alpha_{in}^{\...
...f{T}}}+ 2 \alpha_{null}^{\mathrm{\textsf{T}}}Q_c) \alpha_{in} \right) \; \times$  
    $\displaystyle \qquad \qquad \qquad \qquad \exp \left( - \lambda ( 2 \alpha_{pv,...
... ) \alpha_{null} \right) \; \, d\alpha_{pv,i} \; d\alpha_{in} \; d\alpha_{null}$  

Integrating with respect to $ \alpha_{in}$ gives

    $\displaystyle \int \exp \left( \frac{-\beta}{2} (\alpha_{in}^{\mathrm{\textsf{T...
...+ 2 \alpha_{null}^{\mathrm{\textsf{T}}}Q_c) \alpha_{in} \right) \, d\alpha_{in}$  
    $\displaystyle \qquad \qquad = \int \exp \left( \frac{-\beta}{2} (\alpha_{in}^{\...
... \alpha_{in} + 2 B_0^{\mathrm{\textsf{T}}}\alpha_{in} ) \right) \, d\alpha_{in}$  
    $\displaystyle \qquad \qquad = \left( \frac{\beta}{2} \right)^{-D_{in}/2} \vert ...
...-1/2} \exp \left( \frac{\beta}{2} B_0^{\mathrm{\textsf{T}}}A_0^{-1} B_0 \right)$  

where $ A_0 = G_{in}^{\mathrm{\textsf{T}}}G_{in} + \frac{2 \lambda}{\beta} Q_{in}$ and $ B_0 = G_{in}^{\mathrm{\textsf{T}}}G_{pv,i} \alpha_{pv,i} - G_{in}^{\mathrm{\te...
...2\lambda}{\beta} ( Q_a \alpha_{pv,i} + Q_c^{\mathrm{\textsf{T}}}\alpha_{null} )$.

Substituting this back into the previous expression and integrating over $ \alpha_{null}$ gives

    $\displaystyle \int \exp \left( \frac{-\beta}{2} (Y - G \alpha)^{\mathrm{\textsf...
...Y - G \alpha) - \lambda \alpha^{\mathrm{\textsf{T}}}Q \alpha \right) \, d\alpha$  
    $\displaystyle \quad = \left( \frac{\beta}{2} \right)^{-D_{in}/2} \vert \det(A_0...
...ambda \alpha_{pv,i}^{\mathrm{\textsf{T}}}Q_{pv} \alpha_{pv,i} \right) \; \times$  
    $\displaystyle \qquad \exp \left( \frac{\beta}{2} (Y^{\mathrm{\textsf{T}}}G_{in}...
...xtsf{T}}}G_{pv,i} + \frac{2\lambda}{\beta} Q_a) \alpha_{pv,i} \right) \; \times$  
    $\displaystyle \qquad \exp \left( - \lambda \left[ 2 Y^{\mathrm{\textsf{T}}}G_{i...
...xtsf{T}}}) \right] \alpha_{null} \right) \; \, d\alpha_{pv,i} \; d\alpha_{null}$  
    $\displaystyle \quad = \left( \frac{\beta}{2} \right)^{-D_{in}/2} \vert \det(A_0...
...thrm{\textsf{T}}}\alpha_{null} ) \right) \; \, d\alpha_{pv,i} \; d\alpha_{null}$  
    $\displaystyle \quad = \left( \frac{\beta}{2} \right)^{-D_{in}/2} \vert \det(A_0...
...eft( \lambda B_1^{\mathrm{\textsf{T}}}A_1^{-1} B_1 \right) \; \, d\alpha_{pv,i}$  
    $\displaystyle \quad = \left( \frac{\beta}{2} \right)^{-D_{in}/2} \lambda^{-D_{n...
...{pv,i} + 2 B_3^{\mathrm{\textsf{T}}}\alpha_{pv,i}) \right) \; \, d\alpha_{pv,i}$  

where
$\displaystyle R_0$ $\displaystyle =$ $\displaystyle I - G_{in} A_0^{-1} G_{in}^{\mathrm{\textsf{T}}}$  
$\displaystyle A_0$ $\displaystyle =$ $\displaystyle G_{in}^{\mathrm{\textsf{T}}}G_{in} + \frac{2 \lambda}{\beta} Q_{in}$  
$\displaystyle A_1$ $\displaystyle =$ $\displaystyle Q_{null} - \frac{2\lambda}{\beta} Q_c A_0^{-1} Q_c^{\mathrm{\textsf{T}}}$  
$\displaystyle B_1^{\mathrm{\textsf{T}}}$ $\displaystyle =$ $\displaystyle Y^{\mathrm{\textsf{T}}}G_{in} A_0^{-1} Q_c^{\mathrm{\textsf{T}}}+...
...mbda}{\beta} Q_a^{\mathrm{\textsf{T}}}A_0^{-1} Q_c^{\mathrm{\textsf{T}}}\right)$  
$\displaystyle A_2$ $\displaystyle =$ $\displaystyle G_{pv,i}^{\mathrm{\textsf{T}}}G_{pv,i} + \frac{2\lambda}{\beta} Q...
...}}A_0^{-1} ( G_{in}^{\mathrm{\textsf{T}}}G_{pv,i} + \frac{2\lambda}{\beta} Q_a)$  
$\displaystyle B_2^{\mathrm{\textsf{T}}}$ $\displaystyle =$ $\displaystyle Y^{\mathrm{\textsf{T}}}G_{in} A_0^{-1} (G_{in}^{\mathrm{\textsf{T}}}G_{pv,i} + \frac{2\lambda}{\beta} Q_a) - Y^{\mathrm{\textsf{T}}}G_{pv,i}$  
$\displaystyle A_3$ $\displaystyle =$ $\displaystyle A_2 - \frac{2\lambda}{\beta} \left( Q_b^{\mathrm{\textsf{T}}}- G_...
...hrm{\textsf{T}}}A_0^{-1} Q_c^{\mathrm{\textsf{T}}}\right)^{\mathrm{\textsf{T}}}$  
$\displaystyle B_3^{\mathrm{\textsf{T}}}$ $\displaystyle =$ $\displaystyle B_2^{\mathrm{\textsf{T}}}+ Y^{\mathrm{\textsf{T}}}G_{in} A_0^{-1}...
...mbda}{\beta} Q_a^{\mathrm{\textsf{T}}}A_0^{-1} Q_c^{\mathrm{\textsf{T}}}\right)$  
$\displaystyle R_1$ $\displaystyle =$ $\displaystyle R_0 - \frac{2\lambda}{\beta} G_{in} A_0^{-1} Q_c^{\mathrm{\textsf{T}}}A_1^{-1} Q_c A_0^{-1} G_{in}^{\mathrm{\textsf{T}}}$  

However, as shown above, if all the previous integrations are performed first, then the remaining posterior takes the form

$\displaystyle p(T\vert Y,S,Q,\beta,\lambda)$ $\displaystyle \propto$ $\displaystyle p(T) \, C_1^{D_{un}} \, \vert\det(Q)\vert^{1/2} \vert\det(Q_{null...
...ht)^{(N-D_{in}-D_{un})/2} \left(\frac{\lambda}{\pi}\right)^{D_{in}/2} \, \times$  
    $\displaystyle \qquad \left\vert\det\left( G_{un}^{\mathrm{\textsf{T}}}G_{un} \r...
...f{T}}}R_{un} G_1 + \frac{2\lambda}{\beta} Q'\right)\right\vert^{-1/2} \; \times$  
    $\displaystyle \qquad \qquad \int \exp\left( \frac{-\beta}{2} (Y - G_{pv} \alpha...
...athrm{\textsf{T}}}Q'' \alpha_{pv,i} \right) \, d\alpha_{pv,i} \, d\alpha_{pv,u}$  

where $ R_{un} = I - G_{un} ( G_{un}^{\mathrm{\textsf{T}}}G_{un} )^{-1} G_{un}^{\mathrm{\textsf{T}}}$; $ R = I - R_{un} G_1 \left[G_1^{\mathrm{\textsf{T}}}R_{un} G_1 + \frac{2\lambda}{\beta}
Q'\right]^{-1} G_1^{\mathrm{\textsf{T}}}R_{un}$; $ G_1 = [ \, G_{pv,i} \; \; G_{in} \, ] $; and

$\displaystyle Q' = \left[ \begin{array}{ccc}
Q_{pv} - Q_{b}^{\mathrm{\textsf{T...
...} - Q_{cross}^{\mathrm{\textsf{T}}}Q_{null}^{-1} Q_{cross}
\end{array} \right]
$

and $ Q'' = (Q_{pv} - Q_{b}^{\mathrm{\textsf{T}}}Q_{null}^{-1} Q_{b}) - (Q_{a} - Q_{...
...1} Q_{cross})^{-1} (Q_{a} - Q_{cross}^{\mathrm{\textsf{T}}}Q_{null}^{-1} Q_{b})$.

Assuming that $ G_{pv,u}^{\mathrm{\textsf{T}}}R_{un} R R_{un} G_{pv,u}$ and $ (G_{pv,i}^{\mathrm{\textsf{T}}}R_{un} R R_{un} G_{pv,i} + (2\lambda/\beta) {Q'}_{pv})$ both have negligible off-diagonal terms gives

$\displaystyle p(T\vert Y,S,Q,\beta,\lambda)$ $\displaystyle \propto$ $\displaystyle p(T) \, C_1^{D_{un}+D_{pv,u}} \, \vert\det(Q)\vert^{1/2} \vert\de...
...}-D_{pv,u})/2} \left(\frac{\lambda}{\pi}\right)^{(D_{in}+D_{pv,i})/2} \, \times$  
    $\displaystyle \qquad \left\vert\det\left( G_{un}^{\mathrm{\textsf{T}}}G_{un} \r...
...}}}R_{un} G_{in} + \frac{2\lambda}{\beta} Q'\right)\right\vert^{-1/2} \; \times$  
    $\displaystyle \qquad \exp\left( \frac{-\beta}{2} Y^{\mathrm{\textsf{T}}}R_{un} ...
... q_j^{-1} w_j \right) \, \left( \prod_{j=1}^{D_{pv,i}} {q'}_j^{-1} w'_j \right)$  
  $\displaystyle \propto$ $\displaystyle p(T) \, C_1^N \, \vert\det(Q)\vert^{1/2} \vert\det(Q_{null})\vert...
...v,u})/2} \left( \frac{2 \lambda}{\beta} \right)^{(D_{in}+D_{pv,i})/2} \, \times$  
    $\displaystyle \qquad \left\vert\det\left( G_{un}^{\mathrm{\textsf{T}}}G_{un} \r...
...}}}R_{un} G_{in} + \frac{2\lambda}{\beta} Q'\right)\right\vert^{-1/2} \; \times$  
    $\displaystyle \qquad \exp\left( \frac{-\beta}{2} Y^{\mathrm{\textsf{T}}}R_{un} ...
... q_j^{-1} w_j \right) \, \left( \prod_{j=1}^{D_{pv,i}} {q'}_j^{-1} w'_j \right)$  

where
$ R_{pv,u} = I - R^{1/2} R_{un} G_{pv,u} ( G_{pv,u}^{\mathrm{\textsf{T}}}R_{un} R R_{un} G_{pv,u})^{-1} G_{pv,u}^{\mathrm{\textsf{T}}}R^{1/2} R_{un}$;
$ R_{pv} = I - R_{pv,u} R^{1/2} R_{un} G_{pv,i} ( G_{pv,i}^{\mathrm{\textsf{T}}}...
.../2} R_{un} G_{pv,i})^{-1} G_{pv,i}^{\mathrm{\textsf{T}}}R_{pv,u} R^{1/2} R_{un}$;
$ q_j^2$ is the $ j$th diagonal of $ ( G_{pv,u}^{\mathrm{\textsf{T}}}R_{un} R R_{un} G_{pv,u})$;
$ {q'}_j^2$ is the $ j$th diagonal of $ ( G_{pv,i}^{\mathrm{\textsf{T}}}R_{un} R^{1/2} R_{pv,u} R^{1/2} R_{un} G_{pv,i} + (2\lambda / \beta) Q'')$;

$\displaystyle w_j = \frac{1}{2} - \frac{1}{2} \ensuremath{\mathrm{erfc}}\left( ...
...eta}{2}\right)^{1/2} \left(q_j C_1^{-1} - (R^{1/2} R_{un} Y)_j \right) \right)
$

and

$\displaystyle w'_j = \frac{1}{2} - \frac{1}{2} \ensuremath{\mathrm{erfc}}\left(...
...)^{1/2} \left({q'}_j C_1^{-1} - (R_{pv,u} R^{1/2} R_{un} Y)_j \right) \right).
$


next up previous
Next: Summary Up: Multi-Variate Gaussian Intensity Prior Previous: Marginalisation over Areas of