A Framework for Efficient Detection of Brain Tumor

Abstract

One of the most crucial tasks in any brain tumor detection system is the isolation of abnormal tissues from normal brain tissues. Interestingly, domain of brain tumor analysis has effectively utilized the concepts of medical image processing, particularly on MR images, to automate the core steps, i.e. extraction, segmentation, classification for proximate detection of tumor. Research is more inclined towards MR for its non-invasive imaging properties. Human brain tumor detection and classification are time-consuming however vital tasks for any medical expert. Assistance via computer aided diagnosis or detection systems is commonly used to enhance diagnosis capabilities attaining maximum detection accuracy. But such systems are becoming challenging and are still an open problem due to variability in shapes, areas, and sizes of tumor. The past works of many researchers under medical image processing and soft computing have made noteworthy review analysis on automatic brain tumor detection techniques focusing segmentation as well as classification and their combinations. Despite significant research, brain tumor segmentation is still an open challenge due to variability in image modality, contrast, tumor type, and other factors. Many great works ranging from manual, semiautomatic, or fully automatic tumor segmentation with MR brain images are available, however, still creating a space for developing efficient and accurate approaches in this domain. In this research work, various brain tumor detection techniques for MR images are reviewed along with the strengths and difficulties encountered in each to detect various brain tumor types. The current segmentation, classification, and detection techniques are also conferred emphasizing on the pros and cons of the medical imaging approaches in each modality. The survey presented aims to help researchers to derive essential characteristics of brain tumor types and identifies various segmentation/classification techniques which are successful for detection of a range of brain diseases. An attempt to summarize the current state-of-art with respect to different tumor types would help researchers in exploring future directions. The research proposes a hybrid weighted fuzzy k-means (WFKM) brain tumor segmentation approach using MR images to retrieve more meaningful clusters. It is based on fuzzification of weights which works on spatial context with illumination penalize membership approach which helps in settling issues with pixel’s multiple memberships as well as exponential increase in number of iterations. The segmented image is further utilized for successful tumor type identification as benign or malignant by means of SVM. Experimentation performed on MR images using Digital Imaging and Communications in Medicine (DICOM) dataset shows that fusion of proposed WFKM and SVM outperforms many existing approaches. Further, performance evaluation parameters show that the proposal produces better overall accuracy. Results on variety of images further prove applicability of the proposal in detecting ranges and shapes of brain tumor. The proposed approach excels qualitatively as well as quantitatively reporting an average accuracy of 97% on DICOM dataset with total number of images varying from 100 to 1000. Furthermore, the proposed WFKM in integrated with Hadoop-MapReduce and MDCS platforms to introduce scalability as well as parallelism. The generation of voluminous MR data, however has raised the need to handle massive MR data in significant time. The work thus focuses on analyzing the performance of WFKM in a scalable, distributed, and parallel processing environment. The performance of the collaboration is explored on variable sized datasets i.e. 215MB, 1.2GB and 7.3GB using single node (standalone), 2-node and 3-node configurations. With an increase in the size of MR image data, the processing-time decreases as the number of nodes gets increased, however, authenticate the objective of proposed work. On the other hand, the read and write operation times elevate as the data size increases irrespective to multiple nodes in the configuration. The respective classification accuracy for 215MB, 1.2GB and 7.3GB are reported as 97.07%, 97.01%, and 96.30%. However, reduction in 90% processing-time is observed for 3-node configuration as compared to standalone on a dataset of 215MB. Similarly, 89% and 88.7% decrease in the processing time is noted with 1.2GB and 7.3GB of data, respectively.