Well, I’m in my last year of the Integrated Master in Informatics and Computer Engineering, and I now have to write a thesis. I chose to do a work in the area of medical imaging, using OpenGL, GLSL and CUDA as tools. The context in which this project is inserted has to do with the generation of 3D models of the spine. 3D models of the spine are needed, for example, to evaluate deformities such as scoliosis. There are a number of clinical parameters that can only be measured by the examination of 3D models to correctly assess the condition of the patient and its treatment. To obtain these models what is normally done is a CT or MRI. However, neither of these two methods of obtaining the models are very good:
- they require the patient to be lying down, which may cause changes in the spine configuration
- they are expensive
- they deliver high doses of radiation
All these reasons turn these methods into approaches to avoid during multiple follow ups on the same patient. It’s here that another alternative arises:
- what if there was a way to build the same 3D model of the patient’s spine from two radiographs which capture the front and lateral perspectives of the patient?
If this was possible, it would bring some advantages, namely the cost and amount of radiation delivered. Also, they don’t require the patient to be lying down! The method of recreating the 3D structure of the spine using two radiograph is called biplanar radiography or stereoradiography. Well, I have nothing to do with this method in concrete 
I will be responsible to implement some algorithms to generate Digital Reconstructed Radiograph (DRR).
The thing is that, to make this method work, things need to be fast: the reconstruction of the spine using two radiographs needs to be as fast as possible and to cause as less difficulties as possible to the clinical technicians. As always, it is not enough to prove it is possible: we have to make it usable in the context it is designed for, for the people who will be using it.
So, in the way of making all of this as good as possible, stands the implementation of certain algorithms to produce DRRs from 3D models of bones (vertebrae, more specifically). (The objective is beyond what I want to explain here, but it has to do with comparing the final model of the spine with the two original radiographs…)
These bone models are a mesh of polygons that are to serve as inputs to the algorithms I’m supposed to develop.
We already have an implementation of the algorithm to do this: Single Depth multi-pass peeling with shaders, using OpenGL and GLSL. Ultimately, my thesis’ objective is to implement two new, faster algorithms. One is just an improvement of what is already done, using the same technologies and the second one will be implemented using CUDA, the Nvidia GPU programming language (or extension, or architecture, or whatever…).
If I’m successful, tests and comparisons will be done using these algorithms against what is now the state of the art, and the work done will be integrated in the more broader work which is to develop a process and software capable of producing accurate 3D models of the spine using two radiographs from the patient, thus supplanting the role of CTs or MRIs in diagnosing and treating scoliosis patients.
Well, it’s just about this. Hope it is of interest to some of you
Disclaimer
In general, the above pictures the context of my work. However, it is normal that I’ve just written some more inaccurate parts of the story. Don’t blame me for that
Post a Comment