In simultaneous PET-MRI, attenuation correction is still a major hurdle due to the high attenuation of the bones and the lack of MR signal in conventional sequences. So far, several approaches have been proposed for bone attenuation correction [1-3], including bone segmentation and direct bone imaging. However, almost all available bone segmentation literature focused on the head, which is arguably one of the easier regions because of its smaller field-of-view (FOV) requirement and the absence of major motion artifacts. Direct bone imaging is another promising approach which is accomplished by using zero-TE imaging, but its application in the body is challenging due to the larger FOV requirement and current instrumentation limitation such as peak B1 [2].

Recent research has demonstrated that PET quantitation can be largely improved even by assigning a fixed bone attenuation value (0.120 cm^-1) to all bones [1, 2]. In light of this, we developed a technique that is able to produce good bone segmentation in the pelvic region using a 2-minute 6-echo DIXON MRI acquisition.

DIXON MRI acquisition

All DIXON-MRI scans were acquired on a 3T simultaneous PET-MRI scanner (Siemens Biograph mMR) using a 3D 6-echo GRE sequence with TE optimized for fat water separation (Acquisition time= 2:04, resolution = 2⨉2⨉4 mm3, TR=11ms, bipolar readout, readout direction = LR). The study was approved by local IRB and informed consent was obtained before each acquisition. A graph-cut algorithm was used to generated fat/water images and R2* maps [4] after correcting phase difference associated with bipolar readout.

Segmentation

Bone segmentation was performed with an automated algorithm using the various images generated bythe graph-cut algorithm (denoted as I₁, …, I_n). For each voxel, a multi-dimensional feature vector can be utilized for the purpose of target extraction. The segmentation starts with user drawing some initial sparse seeds $$$P\subset R^3$$$ in the target (bone) region, as well as some rejection seeds $$$Q\subset R^3$$$ in the non-target region. After this, for all the undetermined points in the image, intuitively, each point was assigned to the target or non-target region based on their “distance” to P and Q which is defined by $$$d(x,y) = |I(x)-I(y)|_2$$$ for any two points $$$x$$$ and $$$y\in R^3$$$, where $$$I(x):=(I_1(x), ..., I_n(x))$$$ is a n-dimensional vector storing all the MR information. Such a geodesic calculation process can be implemented using a fast graph algorithm [5].

3D rendering of the segmentation result was performed in Mango (University of Texas Health Science Center).

Figure 1 shows a set of images (fat image, water image, fat water ratio map, R2* map) for a representative slice and the corresponding bone segmentation result. As shown here, the bone segmentation accuracy is reasonably good. Even for regions suffering from artifacts induced by respiratory motion as shown in Figure 2, the segmentation was still acceptable, especially considering the PET resolution is about 4.5mm. Respiratory motion can be reduced by using respiratory gating at the cost of longer acquisition time.

The 3D rendering of the segmentation result of the entire pelvis region is shown in Figure 3. Although the overall segmentation quality appears to be satisfactory, inaccuracies can still be observed in some regions such as the pubic bones and the coccyx.

We plan to apply the bone segmentation technique presented here to bone attenuation map generation for pelvis PET-MRI scans by assigning a fixed bone attenuation value, which in turn is expected to improve the quantitation accuracy of simultaneous PET-MRI.

In this work, our preliminary result demonstrated an algorithm that is able to generate bone segmentation with visually satisfactory accuracy for the purpose of bone attenuation correction. Further study is needed to improve the proposed technique and further reduce user interaction. Its accuracy needs to be established by comparing to a gold standard such as CT or PET transmission images; PET quantitation accuracy with bone attenuation maps generated by this technique also needs to be investigated.