Good afternoon everyone. It's really my great pleasure to present my PhD work today. I would

like to start the talk with a brief introduction of interventional imaging and image guidance

to give you a bit of background on image registration. Here's a picture of an interventional C-arm

arm. This is firstly an X-ray machine with an X-ray detector and a source and detector

mounted on the two ends of the C-shaped arm. Usually the patient is placed in between a

source and detector and under continuous X-ray fluoroscopy. The doctors can navigate their

devices, endovescular devices using these images in real time. It's a standard routine

for endovescular interventions but sometimes it has certain limitations. Firstly, the

vescular structures have very bad contrast in the X-ray images and they are only visible

when the contrast material is injected. Secondly, there's no depth information in such X-ray

images and sometimes it can be quite critical for complex use cases. To accommodate such

limitations, we can use 2D-3D image fusion to augment X-ray fluoroscopy with an overlay

over the 3D volume. If during X-ray imaging we have a 3D volume of the patient and what

we can do is we can virtually place this volume to the patient position and render this volume

as imaged from the X-ray source to detector and now by adding these two images we have

a fused view of both 2D and 3D images. From a technical perspective and for image registration

there are two main factors to consider namely the visualization and accuracy. Visualization

is about how to represent information from both 2D and 3D images and accuracy is about

the spatial alignment between the 3D and 2D images. It leads to the topic of image registration

and in this talk I'm going to focus on the 2D-3D register registration namely by applying

3D rotation translations to the 3D image such that both this 3D image and 2D image are accurately

aligned. This registration workflow consists basically of two parts. Before the interventional

procedure starts this 3D volume has to be brought to the imaging as a projection geometry

according to the patient position and this is performed, this is done by an initial registration.

So now before the procedure the images are well aligned but during the procedure if there's

a patient motion and this patient motion causes misalignment in the fused view and there we

need to apply motion compensation to transform the 3D image accordingly such that the fused

view is again accurate. Of course before I started to work on this topic there was already

a lot of work having been published in this field and there's a very nice review article

summarizing the 2D-3D registration methods and for rigid 2D-3D registration you can categorize

the methods first according for example according to the dimensionality and there we have projection-based,

back-projection-based and reconstruction-based methods and we can also categorize the registration

methods according to the registration basis for the similarity measure so there are then

feature-based, intensity-based and gradient-based. And among all these methods the registration

problem is usually solved as the optimal search on the similarity function S defined under

parameter space and because of the high non-convexity of the optimization problem and that makes

the optimal search very tricky and besides that for intervention applications there are

also special challenges and firstly during the intervention time is a very critical factor

and that means if we have a registration algorithm that delivers very reliable results but if

it takes minutes to calculate the doctors cannot really benefit from it. Secondly content

mismatches can be introduced from the interventional devices such as casters or guide wires or

introduced by contrast material injection and that challenges the robustness of the

registration algorithm a lot. And thirdly a single view guidance is very commonly used

in the practice. It's released less dose and single view system takes also less space but

2D, 3D image registration using single view images is far more challenging than simultaneously

using multiple images from different viewing directions and all these challenges this motivates

me to research the 2D, 3D registration from a different perspective and that leads to

my original contribution to this field. At the beginning I instead of solving I try to

solve the 2D, 3D registration problem is that I assume that an accurate initial alignment

is given and there I can recover the depth information of the 2D features from the fused