D-SegNet

Abstract

Brain tumor is the abnormal growth of cell in brain tissues. These tumors can be benign or malignant based on their position and volume. A pathologist manually inspects the tumor from MRI images, but this process takes time to process and may lead to error in judgment. This leads to automatic detection of the tumor where many significant methods were developed during the decades. Supervised deep learning methods improved the tumor segmentation accuracy considerably due to its ability to detect the tumor pixels. SegNet, an example of deep learning–based tumor segmentation is implemented using Convolutional Neural Network (CNN). Here we propose a modified SegNet called D-SegNet for pixel-wise classification of brain tumors. The method uses dual SegNet layers for extracting the features from MRI images during encoding. The dual layers extract features from images in both horizontal and vertical directions. The conventional square patches are replaced by horizontal and vertical patches to accommodate the complex variation of the tumor pixels. Extracted features are then fine-tuned using singular value decomposition (SVD) and integrated into one for the final decoding process. Different deep learning architectures are implemented to compare performance based on their accuracy. For testing and evaluation, the BRATS2015 dataset is used. From the evaluation, it is understood the proposed method has better accuracy compared to the existing methods.