Automatic Segmentation and Classification of Brain Tumor from MR Images Using DWT-RBFNN

Abstract

Brain tumor segmentation and diagnosis is a very tedious and uncertain task for medical experts for its precise treatment. We have proposed an automatic segmentation and classification of tumor at from brain MR images using DWT-RBFNN classifier. The work is divided into two parts: segmentation and classification. The MR images are 1st preprocessed before applying segmentation and classification on it. In preprocessing stage, MRI images are denoised using hybrid technique DWT-ICA, and resized. The segmentation is performed using hybrid Ostu-canny edge technique and these segmented images are used for the classification of brain tumor. After segmentation stage, 13 types of feature are extracted using multiresolution DWT, these are Median, Variance, Mean, Standard deviation (SD), power spectral density (PSD), RMS (root mean square), Energy, entropy, correlation, homogeneity, skewness, contrast, smoothness. The dataset used in this work contains 443 MRI images of three categories, 105 benign, 223 normal and 105 malignant. The RBNN classifier is used on the segmented and feature extracted images for the classification of brain tumor into three classes, Normal, benign and malignant. To evaluate the performance of the classifier, seven types of evaluation metrics are used, accuracy, F1-score, specificity, classification error rate, recall, precision and overall accuracy. The proposed classifier performance was also compared with Feed forward neural network (FFNN) and Back propagation neural network (BPNN) classifiers. The accuracy result obtained using proposed classifier (RBFNN) is outstanding i.e. 100% on the test dataset as compared to FFNN and BPNN 95.92% and 97.96% respectively.