Abstract:
Brain tumor can grow anywhere in the brain, but it is commonly grown in the cells that make the
brain tissues and the nerves covering the brain’s outer boundary called meninges. There are two
main types of brain tumor known as benign and malignant. Both these types are characterized by
different symptoms and properties, e.g. benign tumor is slow growing as compared to malignant
whereas malignant tumor spreads into the surrounding tissues as opposite to the benign tumor.
The aim of this work is to develop a technique to detect and diagnose the tumor.
This research proposed a brain tumor segmentation method by exploring multi-modality MRI
scans, since brain tumor has unpredictable shape and appearance, which is hard to be captured by
a single modality. The multi-modality images are normally the scans taken by various imaging
modalities e.g. MRI, PET and CT. In our case, we used the images obtained from T1, T2, T1-
Contrast and FLAIR modalities of MRI acquisition. The data from MICCAI BraTS 2013
challenge is utilized, which is co-registered and skull-stripped. Histogram matching and
bounding box is applied to the images of all modalities. Subsequently, the intensity, intensity
differences, local neighbourhood and wavelet features are extracted to develop a system for
identifying the tumor severity.
We further developed an automatic method for segmentation of brain tumor by exploring
machine learning approaches, including random forest, k-nearest neighbour and ensemble
algorithms of adaBoostM2 and rusBoost. We classified voxels into five classes; background,
necrosis, edema, enhancing tumor and non-enhancing tumor and hierarchically computed three
regions (whole tumor, core tumor and enhancing tumor) from above classes.
We applied the above mentioned machine learning approaches and found that random forest
classifier is best among all classifiers tested on our dataset and the extracted features. With leaveone-
out cross validation, it achieved 88% Dice overlap for whole tumor region, 75% for core
tumor and 95% for enhancing tumor, which is better than the Dice overlap reported from
MICCAI BraTS 2013 challenge.
