Ninon Burgos

PhD – Atlas-based methods for image synthesis

While magnetic resonance imaging (MRI) provides high-resolution anatomical information, positron emission tomography (PET) provides functional information. Combined PET/MR scanners are expected to offer a new range of clinical applications but efforts are still necessary to mitigate some limitations of this promising technology. One of the factors limiting the use of PET/MR scanners, especially in the case of neurology studies, is the imperfect attenuation correction, leading to a strong bias of the PET activity. Exploiting the simultaneous acquisition of both modalities, I explored a new family of methods to synthesise X-ray computed tomography (CT) images from MR images. The synthetic images are generated through a multi-atlas information propagation scheme, locally matching the MRI-derived patient’s morphology to a database of MR/CT image pairs, using a local image similarity measure. The proposed algorithm provides a significant improvement in PET reconstruction accuracy when compared with the current correction, allowing an unbiased analysis of the PET images. You can try the method online.

CT synthesis diagram
CT synthesis diagram for a given T1-w MR image. All the MR images in the atlas database are registered to the target MR image. The CTs in the atlas database are then mapped using the same transformation to the target MR image. A local image similarity measure (LIS) between the mapped and target MR images is converted to weights (W) to reconstruct the target CT (pseudo CT).
PET reconstruction results
Example of CT, 18F-FDG PET and 18F-florbetapir (Aβ) PET images obtained from the real CT and the pseudo CTs for an amyloid negative (left) and an amyloid positive (right) subject.

A similar image synthesis scheme was then used to better identify abnormalities in cerebral glucose metabolism measured by 18F-fluorodeoxyglucose (FDG) PET. This framework consists of creating a subject-specific healthy PET model based on the propagation of morphologically-matched PET images, and comparing the subject’s PET image to the model via a Z-score. By accounting for inter-subject morphological differences, the proposed method reduces the variance of the normal population used for comparison in the Z-score, thus increasing the sensitivity.

Subject-specific PET analysis framework
Subject-specific PET analysis framework. The control dataset is first transported into the subject space. To generate the subject-specific Z-score parameters, the patient-specific healthy-population mean (Iμ) and standard deviation (Iσ), the set of registered PETs is locally selected and fused. Finally, the subject-specific Z-map is computed using the model parameters.

To demonstrate that the applicability of the proposed CT synthesis method is not limited to PET/MR attenuation correction, I redesigned the synthesis process to derive tissue attenuation properties from MR images in the head & neck and pelvic regions to facilitate MR-based radiotherapy treatment planning.

Radiotherapy planning example
Example of T2-weighted MR image, reference CT and pseudo CT in the head & neck region.

Main publications

Image synthesis for PET/MR attenuation correction

Image synthesis for the analysis of PET data

Image synthesis for MR-based radiotherapy treatment planning