Development of saliency based algorithm for brain tumor detection

Abstract

Nowadays brain tumor has been deadliest disease across the world. The automatic detection of regions of interest(ROI)is an essential step in the field of medical image processing for proper diagnosis of the tumor to avoid human bias, inter-observer and inter-scanner inconstancy. It will help the radiologists in diagnosing tumor with more accuracy. A novel algorithm has been developed for the identification of tumor regions using visual saliency from magnetic resonance images of the brain. The saliency of an object in an image is the property by means of which it is different from the other objects present within it. A bottom up saliency map is constructed using color distance and spatial distance between windows of the elliptical shape of varying scales. Elliptical windows works to cover curved outliers. As tumor surface also resembles to a curve, we have taken elliptical region of interest in to consideration.As average intensity of each scale is taken as a feature, whenever the window size increases, keeping axis ratio constant, the average intensity value remains unchanged. It makes our detection algorithm to be scale invariant. The computational complexity and time complexity of this detection method is low as we do not have to shift the window across rows and columns to move across the image. Unlike the other state of the art, our algorithmdo does not require any downscaling the input image. It does not need any prior database. The proposed method has been validated on both real and simulated brain images of different patients from MICCAI-BRATS database for the saliency map values. The results obtained are compared with other existing saliency detection methods.The performance evaluation is done quantitatively by using three volume metrics i.e. the Dice similarity index(S),false positive volume function(FPVF),false negative volume function(FNVF) and qualitatively by determining accuracy. The result shows our proposed method is more accurate and robust for tumor detection.