Yaping Wu^{1,2}, Yusong Lin^{2}, Guohua Zhao^{2}, Longfei Li^{2}, and Meiyun Wang^{3}

This study proposes an algorithm to locate and segment Glioma tumor automatically. The algorithm contains three main steps. Firstly, a self-adaptation simple linear iterative clustering (ASLIC0) algorithm was executed to segment T2 weighted MRI images to superpixels images. Then, 52 features including fractal features, curvature feature and higher order derivative map Haralick texture features was calculated on each superpixel. Finally, a Support Vector Machine was trained as a classifier to select superpixels belong to tumor lesion or not. The Dice overlap measure for the segmented Glioma is 0.87 on the data set from the Henan Provincial People’s Hospital.

Introduction

Glioma is a prevalent fatal brain disease, which accounts for approximately 24.7% of all primary brain and other central nervous system tumors, as well as 74.6% of malignant tumors[1]. With the progress of information technology, especially the development of machine learning, the demand for computer-aided diagnosis, adjuvant surgery, radiotherapy and chemotherapy plan and medical research of Glioma patients is more and more intense. Feature calculation is the first step in applying machine learning which means the quantification of the Region Of Interest (ROI) in Glioma tumor. Therefore, the segmentation of ROI is the basis of all subsequent analysis. In practice, existing Glioma segmentation is hard to be applied in clinical routines because of the heterogeneous nature of Glioma and the image acquisition procedures. At present, it is still a challenge work to make an objective, accurate and repeatable brain tumor segmentation. This study proposes an algorithm to locate and segment Glioma tumor automatically.The whole process of Glioma segmentation algorithm presented in this paper was shown in Fig. 1. For the training set, an adaptive superpixel segmentation algorithm ASLIC0 was used to calculate the superpixel maps of Glioma MRI images. Then, fractal features, curvature features and Haralick texture features were calculated for each superpixel. Finally, a SVM classifier was trained to classify the superpixel blocks into tumor region or not, and then the ROI of Glioma is segmented.

The superpixel segmentation algorithm (SLIC) [2] uses K-means clustering algorithm to superpixel processing. The selection of parameters has great influence on the segmentation performance. The parameters of the SLIC algorithm include the number of superpixels K and compactness C. The parameter K determines the number of images divided into irregular regions and he parameter C determines the degree of irregularity. A large number of repeated experiments on MRI images of T2-weighted Gliomas found that there is a direct relationship between the parameter K and the complexity of the region of interest on size and edge, as shown in Fig. 2.

In order to quantify the complexity of the tumor boundary, a complexity index TC was proposed. For Glioma images that have not segmented, the tumor area ratio (TAR) are calculated. This paper selected the K = 200, 120, and 40 respectively according to the difference condition of area ratio and boundary complexity. According to experiment, if TAR and TC are greater than the threshold la, lc, then K selects 200, if TAR and TC are less than the threshold la, lc, then K selects 40, and in other cases K selects 120.

Then, 52 features included Haralick texture features[3], curvature features[4] and fractal features[5] were calculated on each superpixel. These features combined with the training label composed a training set, and then, we classify the tumor area and the non-tumor area by an SVM classifier with linear kernel whose optimization routine was Sequential Minimal Optimization (SMO).

- Q. T. Ostrom et al., "CBTRUS Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2009–2013," Neuro-Oncology, vol. 18, no. suppl_5, pp. v1-v75, 2016.
- R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. Süsstrunk, "SLIC superpixels compared to state-of-the-art superpixel methods," IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 34, no. 11, pp. 2274-2282, 2012.
- R. M. Haralick, K. Shanmugam, and I. H. Dinstein, "Textural Features for Image Classification," Systems Man & Cybernetics IEEE Transactions on, vol. smc-3, no. 6, pp. 610-621, 1973.
- N. Ohtsu, "A Threshold Selection Method from Gray-Level Histograms," IEEE Transactions on Systems Man & Cybernetics, vol. 9, no. 1, pp. 62-66, 1979.
- N. Sarkar and B. B. Chaudhuri, "An efficient differential box-counting approach to compute fractal dimension of image," IEEE Transactions on Systems Man & Cybernetics, vol. 24, no. 1, pp. 115-120, 1994.
- W. R. Crum, O. Camara, and D. L. Hill, "Generalized overlap measures for evaluation and validation in medical image analysis," Medical Imaging IEEE Transactions on, vol. 25, no. 11, pp. 1451-1461, 2006.

Fig.1 Flowchart of
automatic segmentation algorithm

Fig. 2 Influence of tumor area size and boundary complexity on K value.

Note: The first column is the original image, the second column is the manual segmentation, the third column is the superpixel images, the fourth column is the segmentation result by superpixel.

Fig. 3 Glioma image segmentation results.

Note: the first and the fifth columns are the original images, the second and the sixth column are manually defined, the third, the fourth, the seventh, and the eighth columns were segmentation result. The first to third rows were results on low grade gliomas while the fourth to sixth rows were high grade Gliomas.