MR-based radiation therapy (RT) simulation has been receiving considerable attention in the recent decade, owing to the advantage of protecting radiosensitive tissues and superior soft-tissue contrast. One major challenge of MR-simulation is its inefficacy to identify bone which is essential for RT simulation and planning.

Conventional skull segmentation approaches (e.g. [1]) concern absolute intensity information which can be inconsistent in same types of tissues in MR images. It is mainly due to B1/B0 inhomogeneity and imaging artifacts. In particular, B1 inhomogeneity is more pronounced in MR scans for RT simulation purpose because surface coils have to be used in the presence of RT immobilisation devices.

This abstract presented a framework that used local shape prior information to analyze structure shape within local regions, along with global topology prior information to automatically segment the skull. Both prior information was robust to inhomogeneous intensity and required neither training nor case dependent data. The ultimate goal of skull segmentation is to assign electron density information based on the segmented skull to facilitate MR-based RT simulation and planning. In addition, the proposed method is possible to enhance bone-image contrast, perform noise reduction or bone segmentation in various RT-oriented bone imaging sequences.

Four healthy volunteers were immobilised with personally customized thermoplastic mask and scanned on a dedicated 1.5T wide-bore MR simulator (Optima MR450w, GE Healthcare, Milwaukee, WI, US) using surface array coils. We used a low flip angle gradient echo sequence [3] which minimized the intensity contrast among soft tissues (TR/TE=8.6/4.2ms, flip angle=5^o, FOV=24cm, 1.2mm isotropic voxel). A researcher manually segmented the skull of the first volunteer as a reference to quantify automatic segmentation accuracy.

Local skull patches produced much stronger intensity changes along its normal direction than along its tangent plane. At each patch, the second order statistics analyzed by eigen-decomposition [2] yielded one significant principle direction with strong intensity changes, and two principle directions returning minor ones (Fig. 1). This analysis gave large positive eigenvalues inside the skull, small negative values in the vicinity of the skull and insignificant random values the rest of the regions. The local patches detection result (computed as the largest-magnitude eigenvalues of the second order statistics [1]) $$$R(\vec x)$$$ at each position $$$ \vec x $$$ was registered with the manually defined topology template $$$T(\vec x)$$$ to acquire the optimal transformation field $$$F(\vec x)$$$. The topology template is an image encapsulating whether a local patch at $$$\vec x$$$ had lower, higher or unknown intensity compared to the neighbor. It was represented using $$$T(x)=$$$ +1, -1 or 0 respectively (see Fig.2 for the iso-surface of the template). Mathematically, the registration obtained

$$$ arg max$$$_F(x)$$$T(F(x))R(x)$$$.

In our experiments, $$$F(x)$$$ is computed using B-Spline [3]. A diffeomorphic regularization [4] was added to maintain template topology during registration. The segmentation framework was performed using customized MatLab (2014b, Natick, MA, US) scripts.

Automatic segmentation was successful in all subjects (Fig. 3). The dice similarity score was 83.4% for the manually segmented case (the left case in Fig.3). Fig. 4 presented the absolute intensity and image gradient histograms within the segmented skulls and brain regions. The histograms echoed the wide-spread of intensity and gradient, imposing a major barrier on MR-based bone segmentation in existing approaches. Nonetheless, the presented framework well-handled the inhomogeneous intensity and gradient because of the absolute intensity-independent structural and topology priors.

Furthermore, the topology template (Fig. 2) had a shape considerably different from all segmented skulls. It showed that the presented skull segmentation framework tolerated significant shape difference between the topology template and target, thus enabling one topology template to suit different cases.

Separating air cavity from cortical bone is yet challenging due to the indistinguishable MR signal of these two regions. One may consider incorporating the proposed framework in ultra-short TE sequence [5]. On the other hand, CT scans could be obtained to evaluate segmentation accuracy. However, attention will be needed to handle MR geometric distortion when comparing MR and CT images.

In conclusion, a novel automatic skull segmentation method based on local structural prior and global topology prior was proposed. The former was computed using eigen-decomposition on second order intensity statistics to extract local planar patches. The latter was accomplished through registration to maximize the number of patches that a topology template connected, without changing the topology of the template. They jointly yielded an effective framework to perform automatic MR-based skull segmentation. This method is beneficial for MR-based RT planning to reduce ionizing radiation in radiotherapy.