Özet:
Brain tumor is an abnormal and uncontrolled growth of the cells. Early brain tumor
detection is essential to save lives. In fact, brain tumors are difficult to diagnose,
requiring specialized equipment and training. A medical decision making system
facilitates diagnostic process by visualizing the data produced by a classification
system, allowing doctors to make a right diagnosis.
This study proposes an automated system for segmentation and classification the brain
tumor grades in MRI into three classes: normal, LGG and HGG. In the proposed
system, a new segmentation method named LDI-Means algorithm (Local Difference
in Intensity-Means algorithm) is used. It is a clustering technique based on the
difference in the intensity level of one pixel than another. Furthermore, a new
approach in selecting the sub-significant set of attributes is used, denoted MI+SVD
(Mutual Information + Singular Value Decomposition). The robust features are later
used as an input to the classifier. The new network structure called simplified RNN
(Residual Neural Network) is also offered by this study.
The proposed automated system has six stages; the pre-processing, clustering by
LDI-Means, feature extraction, feature selection and dimension reduction by MI+SVD,
and classification by simplified RNN.
The experimental findings at the end of the segmentation stage presented an
approximate match of 99.02% with the hand-labeled images. In addition, in
comparison to the original feature space and two standard dimension reduction
xiv
methods, PCA and SVD, the MI+SVD algorithm offered a more efficient result
for improving the classification process to achieve a satisfied grading of brain
tumors. Furthermore, using a simplified RNN as a classifier provides a high level
of effectiveness to the proposed system. In comparison with other published studies,
it is found that the proposed system is very sufficient to offer a meaningful real-time
estimation for identification the brain tumor grades.