Bharath K. Navalpakkam1,
Harald Braun2, Joachim Hornegger1, Torsten Kuwert3,
Harald H. Quick2
1Pattern
Recognition Lab, Friedrich-Alexander-University Erlangen-Nrnberg, Erlangen,
Bavaria, Germany; 2Institute of Medical Physics,
Friedrich-Alexander-University Erlangen-Nrnberg, Erlangen, Bavaria, Germany;
3Clinic of Nuclear Medicine, Friedrich-Alexander-University
Erlangen-Nrnberg, Erlangen, Bavaria, Germany
An MR-based attenuation correction approach for PET/MR that includes cortical bone information is proposed. An epsilon insensitive Support Vector Regression model is trained and validated on head regions for five patients. The reconstructed PET data is evaluated for the proposed approach (PETMRAC) against other approaches that include (PETboneAC) and disregard bone class (PETnoboneAC). For the complete brain, PETMRAC yielded a mean absolute error of 2.403.59%, 10.153.31% for PETnoboneAC and 3.963.71% for PETboneAC. For atlas segmented landmarks, the errors for the same were 2.161.77%, 11.032.26% and 4.223.91%, respectively. An initial result of an MR-based pseudo-CT for a whole-body application is illustrated.