Brain tumor segmentation in multimodal MRI images using novel LSIS operator and deep learning

Determination of tumor extent is the foremost challenge in the brain tumor treatment planning and valuation. Among various conventional anatomical imaging techniques for brain tumor diagnosis, MRI (Magnetic Resonance Imaging) provides the best spatial resolution and is noninvasive. MRI volume study for brain tumor extent segmentation is a time-consuming task. The performance highly relied on radiologist's experience. Also the difficulty of MRI and greater volume of test per radiologist in tumor screening system leads to a significant error. The sensible alternative to this problem is to use fully automated Computer Aided Diagnosis (CAD) system which attains the better segmentation of tumor extent with probable image processing tasks. In this paper we propose novel method that segment tumor from 3D MRI data contains Higher Grade Glioma (HGG). This approach not only localizes the tumor but also segments the intra tumor regions (necrosis, edema, non-enhancing tumor, and enhancing tumor). The projected cascaded Convolutional Neural Networks (CNN) actually has two subnetworks such as Tumor Localization Network (TLN) and LSIS (Local Symmetry Inter Sign) based Intra tumor Segmentation Network (LITSN). In TLN, 3DUNet was used to localize the tumor. Then intra tumor regions are segmented using deep CNN with the proposed novel operator which is based on LSIS. The proposed method was validated on BRATS 2015 datasets which contains high-grade glioma (HGG). Experimental results shown that our method can obtain superior segmentation results compared with other promising methods.

Automated summary

This study presents a novel method for tumor segmentation in brain tumor diagnosis using MRI data and deep convolutional neural networks. The proposed method can accurately localize and segment tumors as well as their intra-tumor regions, thus improving the effectiveness of brain tumor treatment planning.