A Novel Approach to Enhance Effectiveness of Image Segmentation Techniques on Extremely Noisy Medical Images

Abstract

AbstractThrough this study, I contribute towards segmentation of liver areas and have proposed additional improvements, which positively influence image segmentation. In this study, I have subjected medical images from LiTS - Liver Tumour Segmentation Challenge, which are extremely noisy, to various image segmentation techniques belonging to fully automatic and semi-automatic categories. These varied techniques implement different approaches towards image segmentation problem. All the techniques had initially failed to segment the images with very poor results. Commonly used filters for pre-processing, such as median filter, top hat filter, wiener filter, etc., were ineffective in reducing the noise effectively. Through this study, I have introduced a new combinatorial approach which not only is easier to implement but also much faster as well and resulted in much more enhanced input image quality that significantly improved the segmentation outcomes. Our approach has reduced noise, sharpened the edges, “localized” the segmentation problem before subjecting to various segmentation techniques. The techniques which had failed previously now could segment the images with improved speed of execution, efficiency and accuracy. I have studied our approach on 10 well known image segmentation techniques. Accuracy of these segmentation techniques was determined by computing Jaccard Index, Dice Coefficient and Hausdorff Distance. KeywordsNoise removalSegmentation accuracyImage filteringEffectivenessDe-convolution