Multistage Classification and Segmentation of Brain MR Image Using Modified Soft Computing Techniques

Abstract

Recent studies indicate that brain tumor is one of the major causes of human casualties. Timely and accurate diagnosis of this life taking disease could reduce the casualty rate and extend the life of a person. In this research paper, techniques for brain tumor detection from MR Images with malignancy using modified soft computing approaches are presented and analyzed. An automated tumor detection system using artificial neural network (ANN) is proposed to classify the images as any of the four classes: Glioblastoma multiforme, Meningioma, secondary tumor-Metastasis and No Tumor. The classified image undergoes a segmentation process that predicts the size of the tumor in terms of pixels. Traditionally, conventional self-organizing map (CSOM) and Conventional back Propagation network (CBPN) are used for classification and segmentation respectively. However, these methods provide less accurate results in addition to high computational complexity. Moreover, due to unstable target weights, the number of iterations is large. These drawbacks are overcome in the proposed technique by developing a modified SOM (MSOM) for classification of images and modified BPN (MBPN) for segmentation. Simulated results show that the proposed modifications minimize the computational complexity without compromising on the accuracy. It is shown that MSOM increases the accuracy of classification by 10% compared to its conventional counterpart. Similarly segmentation accuracy is improved by 8% using MBPN.