The typical simplifiled representation of a tomographic system is

$y = Ax$,

where $y$ is the collected data (sinogram in CT), $A$ is the projection matrix, and $x$ is the unknown image.

The closed form solution, which is probably the most popular one, is to applying FBP algorithm to get the direct inverse of this model. In other words, $x$ can be solved directly from $y$ using FBP algorithm.

The reason why we don't do $x = A^*y$ is that the the $A^*$ is too large to be computationally possible on a computer. Therefore, FBP is the way to go.

Now I am wondering is there a way to do this reconstruction using machine learning/deep learning? I am not an expert on machine learning topic, but I will present the problem using a simple example.

Basically, what I like is to have an image $x$ when a sinogram $y$ is given. Suppose I have an image (this is my $x$, or output in this case) of size $128 \times 128$. I can generate its sinogram (this is my $y$, or input data in this case) by simulating the physical process (Radon transform). The size of the sinogram should be $128 \times 360$. Which means the beam cast 128 rays across all 360 degrees.

I can generate the sinogram for arbitrary images using my simulator. So I should have a lot of training data.

I imaging between the input $x$ and its output $y$, there is black box in between inside which there are a lot of neurons, weights, bias structured within layers. Now I am wondering how to construct such a neural networks (ex. how many layers? how many weights for each layer?)? and how to train the weights and bias such that when I have a sinogram $y$, it can solve for the image $x$.

If there is a paper talking about this, I'd like to read it. I hope the paper is simple, and an easy tutorial. I am really a beginner on this topic.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.