E-fuzzy feature fusion and thresholding for morphology segmentation of brain MRI modalities

Abstract

Brain tumor segmentation is a significant procedure in medical image processing. Effective and efficient segmentation is always a key concern for the radiologists due to the presence of low illumination in imaging modalities of Magnetic Resonance (MR) imaging. Some of the challenging issues addressed in this paper are the sensitivity of noise, variations, and non-standardization in inter-slice intensity, and intensity inhomogeneity. A reliable segmentation method for the brain tumor segmentation is necessary for efficient measurement of the tumor. The foremost objective of the research is to fuse the different modalities of MRI images by considering the most useful features to obtain the best segmentation. Sufficient diagnostic information in clinical applications is not provided by single modality images. Therefore, combining the features of different modalities of images is vital. Recently, many researchers applied many techniques for fusing medical images, still many issues are to be addressed. Hence, to combine different modality of images, a novel fusion method is proposed. Our proposed work contains four stages (i) pre-processing, (ii) feature extraction (iii) feature fusion, and (iv) segmentation. In the pre-processing step, the image quality is increased and the unwanted noise is removed using average filtering. In the feature extraction stage, the Extended Grey Level Co-occurrence Matric (E-GLCM) is introduced to extract the suitable intensity and texture features of brain tumor MRI images. To fuse the correct features for efficient segmentation, the Enhanced Fuzzy Radial Basis function Neural Network (E-FRBNN) incorporates five layer inputs, fuzzy partition, front combination, inference, and output. Finally, the segmentation is carried by thresholding based segmentation approach with morphology operators. The simulation results of the proposed segmentation procedure acquire competitive performance when compared with the existing techniques for the BRATS 2015 dataset.