Image reconstruction

Medical imaging reconstruction methods have reached remarkable levels of performance, and are now widely used. Today’s challenge is the accumulation of X- and gamma-ray doses and the quest to reduce them, which has a negative impact on system signal-to-noise ratios and, as a result, the diagnostic quality of the images obtained or doses measured.

Furthermore, in the fields of radiotherapy and medical imaging, researchers are often faced with the task of having to determine a quantity of interest, based solely on experimental data, which cannot be directly observed. Problems reconstructing tomographic images and evaluating phase space using radiotherapy dose deposition measurements are just two examples.

Statistical modeling for imaging

R&D scientists working at Doseo have developed solid know-how in statistical modelling, especially in the field of Bayesian nonparametric inference for data analysis and signal and image processing. In this innovative statistical model, the inverse problem is approached as a statistical inference problem in which the quantity of interest is treated as a random measure. On the one hand, Bayesian modelling allows for the regularization of the ill-posed problem. On the other hand, nonparametric calculation provides a flexible and robust inference framework including uncertainty estimation.

Image reconstruction

Our know-how in image reconstruction is useful in addressing various problems encountered in radiotherapy and medical imaging:

• For PET (positron emission tomography) scans, to substantially reduce the doses injected without compromising image quality and, with an estimation of functional volumes and the associated error, objectively assess the therapeutic response during patient treatment monitoring
• For radiotherapy, to develop an innovative denoising algorithm for 2D and 3D dose distributions obtained using Monte Carlo simulations or to develop nonparametric virtual source models to model a beam combining the results of Monte Carlo simulations and the measurements

Our research covers

• Development of image-processing algorithms for multimode imaging (MRI, PET, CT)
• Development of virtual sources combining data from Monte Carlo simulations and Bayesian nonparametric inference

Image reconstruction has several PET scan applications

The use of statistical methods (Bayesian nonparametric inference) led to the development of an image reconstruction algorithm for PET scan images (3D and 3D+time).

The development of a 4D PET statistical image reconstruction method resulted in:

• The reconstruction of images acquired with a lower (by a factor of at least ten) injected dose of radioactivity without compromising on image interpretation quality
• An estimation of functional volumes and the associated uncertainty; this information is used to objectively evaluate the therapeutic response during patient treatment monitoring, for example

This advance was validated on two cameras for clinical data for images corresponding to a dose reduced by a factor of ten while maintaining time resolution better than one second.