Automated 3D segmentation of brain tumor using visual saliency

There is a growing availability of medical imaging data from large number of patients, involving visual information from images in different modalities along with an associated complexity of the features of interest. It has therefore become essential to develop automated delineations to assist doctors and/or radiologists to analyze and speedup medical image understanding, preferably avoiding any user intervention. We present here a novel approach for the reliable, automated, and accurate 3D segmentation of brain tumors from multi-sequence magnetic resonance images. The tumor volume, detected using visual saliency, is evaluated in three-dimensions for small as well as large ROIs and/or VOIs. The proposed segmentation method is applied on the publicly available standard BRATS data set, and is found to achieve very high accuracy with good reliability (or repeatability) and robustness of results. Its robustness, is also investigated by measuring the impact of tumor size on segmentation accuracy, on the basis of the weak linear correlation. The results demonstrate that the segmentation generated by the proposed algorithm can be used for accurate, stable contouring, for both high- and low-grade tumors, as compared to several related state-of-the-art methods involving semi-automatic and supervised learning.