Thank you for the introduction.

So finally I don't need to talk about limited angle tomography today.

So today I will talk about data truncation using deep learning.

So here is the PRS slide.

So you know I have been in the lab for so long.

And since last PRS I have one journal accepted and also one BVM conference.

And also dissertation.

I just received the email.

It's online now.

So okay cool.

I should add it.

Okay so let me introduce data truncation in computed tomography first.

So the problem of data truncation in computed tomography will arise in two scenarios.

One is that in some certain clinical applications, so only a certain region of the patient is

of interest.

For example if we want to deploy a stand to a certain artery or if we want to get some

tissue samples with biopsy.

An X-ray collimator will be placed between the X-ray source and the detector to reduce

dose.

So this is called ROI imaging or also called interior tomography.

So here is an example.

So this is the head projection.

So if we are only interested in the area near the ear, so for example this area, then with

the collimators, then only this part or this region is acquired.

Then we have data truncation laterally and also in this case vertically.

The other scenario is that due to the limited size of flat panel detectors, so they are

not big enough to cover the whole imaged body.

So usually human heads are not so big so it is fine to cover the whole head.

However for abdomen and also for torsos because they have a much bigger size, then the detectors

are not large enough to cover.

Then for example this is an ideal projection and this is in practice with a limited size

projection.

Then we only get the projections in this ROI box or in this detector.

Then the projections are also laterally truncated.

Image reconstruction from truncated data with the classic FBP reconstruction will suffer

from artifacts.

So here this is the reference image and this is the FBP reconstruction.

We can clearly observe this FOV boundary.

Then we can see that the anatomical structures outside the FOV are missing and the intensity

values inside the FOV are much larger than the reference.

Here we can see they are much brighter.

So we call this capping artifact because if we plot a live profile then we can see its

intensities look like a cup.

That's why we call them capping artifacts.

There are many algorithms proposed for data truncation including heuristic extrapolation,

analytic reconstruction, compressed sensing.

And today we will apply deep learning for data truncation.

So far I only know three papers in this application.

One is that from the CAST group they have proposed to apply the UNAT to post-process

FBP reconstructions or DBP reconstructions.