Tomographic reconstruction for low-dose x-ray systems using deep learning

X-ray tomography equipment makes it possible to obtain quality tomographic images but with the main drawbacks of high radiation dose for the patient, avoiding its indication in pediatric applications or for repetitive or screening tests, and the fact that the patient must be completely positioned inside the scanner gantry, which makes its use impossible in cases where there are patient mobility difficulties (in the ICU or during surgery, for example). The recent technical solutions for obtaining tomographic images from limited data (angular coverage of less than 360 degrees) enables the use of these technologies in situations where it is difficult to have a CT scanner available (in an ambulance, for example).

Image reconstruction from limited angle data with a reduced number of projections and not a standard tomographic geometry, requires the development of new techniques to compensate for the lack of data. One possible approach involves the integration of Deep Learning technologies. The main disadvantage of iterative reconstruction algorithms is the computational time, which is a limitation for their use in this type of environment where it is crucial to maintain the real-time condition, so that decisions can be made at the same moment in the operating room or in the ambulance. For these reasons, it is necessary to explore the possibilities of developing these algorithms on non-traditional computing platforms with models and programming paradigms that can provide the necessary computational resources to obtain these images in times shorter than those obtained in traditional systems.

The ultimate goal of this project is to investigate advanced reconstruction methods for limited data using machine learning based techniques and their implementation using GPU based optimization strategies.