Medulloblastoma is a primary, malignant, high grade brain tumor. It is the most common brain tumor in children. About 20% of childhood brain tumors are medulloblastomas. The patients with this disease have to be treated by maximum tumor surgical removal [1]. Accurate segmentation is necessary for this surgery. However, Manual segmentation of medulloblastoma from magnetic resonance images is a time-consuming task. We proposed a novel method for the segmentation of medulloblastoma with minimal user interaction.

Medulloblastoma can be seen clearly in T2 image. It shows heterogeneous due to calcification, necrosis and cyst formation. Overall tumor region is isointensity with grey matter.

Selective Gaussians Filtering Regularized Level Set (SBGRLS) method is first proposed by Zhang [2]. It combines the merits of Chan-Vese model and Geodesic Active Contour model (GAC), which utilizes the global intensity information to construct a singed pressure force (SPF). It modulates the signs of the pressure forces inside and outside the region of interest so that the contour shrinks when outside the object, or expands when inside the contour. It has advantage to efficiently stop the contours at weak or blurred edges, and does not require a prior. However, it is hard to deal with inhomogeneous gray intensity and sensitive to the initial contour. In our modified SBGFRLS, we modified the SPF function by using a kernel of intensity information to overcome the inhomogeneous problem. With addition of brain mask into SPF function, the SPF function can ignore non-brain structures which help increasing the flexibility in ROI setting. The modified level set equation ($$$\phi$$$) is defined as: $$ \frac{\partial \phi}{\partial t} = spf(I'(x))\cdot\alpha|\triangledown\phi|$$ where I’(x) is local mean intensity inside a small kernel and $$$\alpha $$$ is constant propulsion term. New spf function is defined as: $$ spf(I'(x))=\frac{I'(x)-\frac{c1+c2}{2}}{max(I'(x)-\frac{c1+c2}{2})}$$ c1 is the mean intensity value inside brain mask inside contour boundary and c2 is the mean intensity value inside brain mask outside contour boundary, and is defined as: $$c1=\int_{\omega}^{} \frac{I(x)\cdot H(\phi)\cdot B(x) dx}{H(\phi)\cdot B(x) dx} $$ and $$c2=\int_{\omega}^{} \frac{I(x)\cdot (1-H(\phi))\cdot B(x) dx}{(1-H(\phi))\cdot B(x) dx}$$, where B(x) is binary brain mask and H($$$\phi$$$) is aHeaviside function.

A patient with medulloblastoma was selected to be study cohort. T2 image was acquired by 1.5T MRI (Ingenia, Philips Medical Systems) with the following parameters: TR=6046ms, TE=100ms, dimension=672x672x30, spacing= 0.342x0.342x5.5. T2 image is obtained in DICOM format.

Our method consists of five main parts. (1) Brain extraction was first done on the T2 image using BET module of FMRIB Software Library (FSL, Analysis Group, FMRIB, Oxford, UK), which show the most promising result. (2) ROI with tumor was manually cropped to reduce the computational cost. (3) Then, locally ultra-hyperintensity value was normalized in reference with cropped image. (4) The preprocessed image was then segmented by the modified SBGFRLS method. The initial contour is set automatically in the middle of the ROI. (5) Final segmentation was obtained by applying hole filling operation on the binary image after the SBGRLS algorithm as some regions with inhomogenous intensity will be too big for the algorithm to be segmented out.

In our experiment, we choose propulsion term as 20 and kernel size as 3x3x3. The contour converges in 30 iteration. The result segmented by our method agrees well with the result of manual segmentation (ground truth). Our method yield a segmentation result with dice similarity coefficient of 92%.

Computation time for our proposed method is between 60 and 100s on a single CPU running at 2.70 GHz. Computation time mainly depends on the size of image dataset and the region of interest. Comparing with SBGFRLS algorithm and our proposed method, our proposed give more accurate result. As SPGFRLS depends on the difference in global average mean intensity between inside and outside of contour. With improper initialization, SBGFRLS algorithm is easily trapped into the ultra-hyper density region while our method does not have this initialization problem, no matter initialize on the region of necrosis, calcification or cyst formation. Our method is more robust in contour initialization. Moreover, our method gives a smoother edge in the boundary of the tumor.

We presented that our method perform well to segment medulloblastoma with definition of ROI only. This algorithm is a clinical-oriented method which is fast, robust, accurate and little user interaction to segment medulloblastom. This method can still be improved to further segment the region of necrosis, calcification and cyst formation.