Accurate electron density information is extremely important in positron emission tomography (PET) attenuation correction (AC) and radiotherapy (RT) treatment planning (RTP), especially in the head region, as many interesting brain regions are located near the skull. Achieving good electron density information for bone is not trivial when magnetic resonance imaging (MRI) is used as a source for the anatomical structures of the head, since many MRI sequences show bone in a similar fashion as air. Various atlas-based, emission-based, and segmentation-based methods have been explored to address this problem. In this PhD project, a pipeline for MRI-based substitute CT (sCT) creation is developed and novel ways are developed to further improve the quality of bone delineation in the head region.

First, a robust sCT pipeline is developed and validated. This allows modular improvements of the various aspects of head sCT in later publications. The MRI image is segmented into different tissue classes and the final sCT image is constructed from these. The sCT images had good image quality with small non-systematic error. The time-of-flight (TOF) information improves the accuracy of PET reconstruction. The effect of TOF with different AC maps is evaluated to substantiate the need for accurate AC maps for a TOF capable system. The evaluation is performed on both subject and brain region level. While TOF information is helpful, it cannot negate the effect of the AC map quality.

The sinus region is problematic in MRI-based sCT creation, as it is easily segmented as bone. Two new methods for addressing AC in the sinus region are presented. One method tries to find the cuboid that covers the largest area of air tissue incorrectly assigned as bone and then correct the incorrect attenuation coefficient. Another method uses the sinus covering cuboid in the normalized space, from which it is converted back to each subject’s individual space, after which the attenuation coefficients are calculated. Both methods improve the alignment of sCT and CT images. Finally, the possibilities of improving the quality of the bone segmentation by utilizing a random forest (RF) machine learning process is explored. The RF model is used to estimate the bone likelihood. The likelihood is then used to enhance the bone segmentation and to model the attenuation coefficient. The machine learning model improves the bone segmentation and reduces the error between sCT and CT images.