Three MRI-based attenuation correction methods for PET

Objectives PET/MRI has garnered interest as a clinical and research tool. However, MRI-based attenuation correction (AC) is necessary if the PET is to be quantitative. By and large, there are three ways to convert an MRI to an attenuation map (µ-map). 1) Segment the MRI into tissue types and assign each an attenuation coefficient (µ-coefficient), 2) register a template µ-map to the MRI, and 3) construct a function that maps MRI voxels to µ-coefficients. In this work, we compare three MRI-based AC algorithms, one for each approach, by assessing their performance on whole-body PET.

Methods Twelve oncologic patients were imaged with FDG PET/CT and 3T MRI. The MRI was converted into three µ-maps: one by segmentation (air, lung, soft tissue, bone), one by registration (the CT of a gender and BMI matched patient was chosen as the template µ-map), and one by mapping (the function was a trained support vector machine with a radial basis kernel). The PET reconstructions guided by these µ-maps were analyzed in a tissue specific fashion (lungs, fat, muscle and organs, bone). In particular, we computed the accuracy, precision, and root-mean-square error (RMSE) in each tissue using the PET/CT scan as the silver-standard. Additionally, for the segmentation approach, we assessed the impact of ignoring bone.

Results Segmentation ignoring bone yielded the lowest RMSE in the lungs, fat, muscle, and organs. However, it had the worst accuracy and RMSE in bone. In lung and fat, registration was the most accurate but the least precise. Mapping was the most precise in the lungs, while mapping and segmentation were equally precise in fat. All approaches had similar accuracy and precision in muscles and organs.

Conclusions The segmentation approach ignoring bone performed the best. However, it is not suitable when quantification in the bones is of interest. Further, both the mapping and registration approach had some desirable traits in terms of accuracy and precision. A combination of the three methods may prove synergistic.

Research Support Funding provided by the Natural Sciences and Engineering Research Council of Canada, the Canadian Foundation for Innovation, the Canadian Institutes of Health Research, the Ontario Research Fund, and Multi-Magnetics Incorporated