Intensity Inhomogeneity Correction for Magnetic Resonance Imaging of Automatic Brain Tumor Segmentation

Abstract

Automatic segmentation of brain tumor data is a very important task for all medical image processing applications, especially in the diagnosis of cancer. This work deals with some of the challenging issues such as noise sensitivity, partial volume averaging, intensity inhomogeneity, inter-slice intensity variations, and intensity non-standardization. To deal with the above tasks, this work uses the 3D convolutional neural network (3DCNN) for automatic segmentation and a novel N3T-spline intensity inhomogeneity correction for bias field correction. The proposed work consists of four levels: (i) preprocessing, (ii) feature extraction, (iii) automatic segmentation, and (iv) postprocessing. In the first stage, a novel N3T-spline is suggested to correct the bias field distortion for reducing the noises and intensity variations. For the extraction of texture patches, the extended gray level co-occurrence matrix-based feature extraction is used. Then, the proposed 3D convolution neural network automatically segments the brain tumor and divides the various abnormal tissues. Finally, a simple threshold scheme is applied to the segmented results for correcting the false labels and to eliminate the 3D connected small regions. The simulation results in the proposed segmentation approach could attain competitive performance as compared with the existing approaches for the BRATS 2015 dataset.