9.A. Radon transform

$\renewcommand{\Re}{\operatorname{Re}}$ $\renewcommand{\Im}{\operatorname{Im}}$ $\newcommand{\erf}{\operatorname{erf}}$ $\newcommand{\dag}{\dagger}$ $\newcommand{\const}{\mathrm{const}}$ $\newcommand{\arcsinh}{\operatorname{arcsinh}}$

9.A. Radon transform


  1. Definition
  2. Application to wave equation
  3. Application to medical tomography

Definition

\section{Radon transform}

\subsection*{9.A.1 Definition}

We decompose function on $\mathbb{R}^n$ into monochromatic plane waves: \begin{gather*} u(\boldsymbol{x})=\iiint \hat{u}(\xi) e^{i\boldsymbol{x}\cdot \xi}\,d\xi, \end{gather*} where \begin{gather*} \hat{u}(\xi,t)=(2\pi)^{-n}\iiint u(\boldsymbol{x},t) e^{-i\boldsymbol{x}\cdot \xi}\,dx. \end{gather*}

Introducing spherical coordinates in $\mathbb{R}^n\ni \xi$, $\xi=\rho \omega$ with $\rho=|\xi|$ and $\omega\in \mathbb{S}^{n-1}$, which is $(n-1)$-dimensional sphere of radious $1$ in $\mathbb{R}^n$, we see that for odd $n$ \begin{equation} u(\boldsymbol{x})=\frac{1}{2}\int_{-\infty}^\infty \iint_{ \mathbb{S}^{n-1}} \hat{u}(\rho \omega)e^{i\rho \boldsymbol{x}\cdot \omega} \rho ^{n-1}\,d\omega , \label{eq-9.A.1} \end{equation} where we replaced $\int_0^\infty \,d\rho$ by $\frac{1}{2}\int_{-\infty}^\infty \,d\rho$, because $(\rho,\omega)\mapsto (-\rho,-\omega)$ does not change the integrand.

Then \begin{equation} u(\boldsymbol{x})=\iint_{\mathbb{S}^{n-1}} v(\boldsymbol{x}\cdot\omega,\omega)\,d\omega, \label{eq-9.A.2} \end{equation} with \begin{multline} v(s,\omega)=\frac{1}{2}\int_{-\infty}^{\infty} \hat{u}(\rho\omega)e^{i\rho s}\rho^{n-1}\,d\rho=\\ \frac{1}{2}(2\pi)^{-n} \iiint \Bigl(\int \rho^{n-1} e^{-i\rho \boldsymbol{x}\cdot\omega+i\rho s} \,d\rho\Bigr)u(\boldsymbol{x})\,dx=\\ \frac{1}{2}(2\pi)^{1-n}(-i\partial_s)^{n-1} \iiint \delta(x\cdot\omega -s) u(\boldsymbol{x})\,dx=\\ \frac{1}{2}(2\pi)^{1-n}(-i\partial_s)^{n-1} \iint_{\boldsymbol{x}\cdot \omega =s } u(\boldsymbol{x})\,d\Sigma, \label{eq-9.A.3} \end{multline} where $d\Sigma$ is an area element on the $(n-1)$-dimensional plane $\{x\colon x\cdot\omega=s \}$. Here we used that \begin{gather*} \int e^{-i(s'-s)\rho }\,d\rho=2\pi \delta (s-s'). \end{gather*}

Definition 9.A.1 \begin{gather} (Ru)(s ; \omega) :=\iint_{\boldsymbol{x}\cdot \omega =s } u(\boldsymbol{x})\,d\Sigma \label{eq-9.A.4} \end{gather} with $s\in \mathbb{R}$ and $\omega\in \mathbb{S}^{n-1}$ is a \emph{Radon transform} of $u(x)$.

Then \begin{gather} v(s,\omega)= \frac{1}{2}(2\pi)^{1-n}(-i\partial_s)^{n-1} (Ru)(s ; \omega). \label{eq-9.A.5} \end{gather}

Formulae (\ref{eq-9.A.2}) and (\ref{eq-9.A.4}) provide the inverse Radon transform.

Application to wave equation

Let $u(\boldsymbol{x},t)$ satisfy $n$-dimensional wave equation \begin{gather} u_{tt}-c^2\Delta u =0. \label{eq-9.A.6} \end{gather}

Then its Radon transform $(Ru)(s,\omega,t)$ satisfies $1$-dimensional wave equation \begin{gather} (Ru)_{tt}-c^2(Ru)_{ss} =0. \label{eq-9.A.7} \end{gather} Therefore, if $u(\boldsymbol{x},0)=0$ \begin{gather} (Ru)(s,\omega,t)= \frac{1}{2c}\int_{s-ct}^{s+ct} (R\phi)(s',\omega)\,ds',\qquad \phi (x)=u_t(x,0). \label{eq-9.A.8} \end{gather} Let $n=3$, define $v(s,\omega)$ by (\ref{eq-9.A.3}): \begin{multline*} v(s,\omega,t) = -\frac{1}{2}(2\pi)^{-2}(\partial_s)^{2} (Ru)(s, \omega,t)\\ =-\frac{1}{16c}\partial_s \Bigl( (R\phi)(s+ct,\omega) - (R\phi)(s-ct,\omega)\Bigr)\\ =-\frac{1}{16c^2\pi^2}\partial_t \Bigl( (R\phi)(s+ct,\omega) + (R\phi)(s-ct,\omega)\Bigr) \end{multline*} and according to (\ref{eq-9.A.2}) \begin{gather*} u(\boldsymbol{x},t)=-\frac{1}{16c^2\pi^2}\partial_t \iint_{\mathbb{S}^2} \Bigl( (R\phi)(\boldsymbol{x}\cdot\omega+ct,\omega) + (R\phi)(\boldsymbol{x}\cdot\omega-ct,\omega)\Bigr)\,d\omega. \end{gather*} From this one can derive (9.1.4).

Application to medical tomography

The CT-scan is based on the following physical law: in homogeneous media intensity of the passed (refracted) electromagnetic ray decays exponentially \begin{gather*} I = I_0 e^{-\lambda s} \end{gather*} where $I_0$ is the intensity of the original ray and $\lambda$ is the linear attenuation coefficient often informally called optical density along the ray. In inhomogeneous media \begin{gather*} I = I_0 e^{-\int \lambda(s)\,ds} \end{gather*}

Remark

  1. $\lambda$ depends also on frequency but in X-rays frequency is fixed so $\lambda$ depends on position only.
  2. In inhomogeneous media rays are not straight lines but in X-ray practice deviation from the straight lines is negligible.

A larger attenuation (i.e., larger $\lambda$) produces lower transmitted intensity $I$, which results in a brighter region on a traditional X-ray film (radiographic film), because the film behaves as a \emph{negative} (reversing darkness). Modern systems use detector plates instead of film, but the mapping from intensity to displayed brightness still follows the same negative-like convention.

Thus, bones—which attenuate X-rays strongly—appear white or light gray, while soft tissues appear dark gray or black.

Remark

  1. In classical photographic development, film was produced first and then printed onto photographic paper. Both standard film and paper were negatives: the film reversed the brightness, and the final print reversed it back.
  2. Positive films and papers existed but were invented later, were rarer, and more expensive.

CT-scan (CT=Computer Tomography) a.k.a. CAT-scan (CAT=Computer Assisted Tomography) works by "illuminating” the body from many directions perpendicular to a chosen axis (say the $x$-axis) and recording, for each angle, the corresponding Radon transform in the $(y,z)$ plane. That is: we have $2$-dimensional Radon transform, $x$ is just a parameter.

By inverting this transform, one reconstructs $\lambda(x; y, z)$, the attenuation coefficient at each point, producing a 3D internal map of the body.

Read more: Radon transform and CT Scan.


$\Leftarrow$  $\Uparrow$  $\Rightarrow$