Tissue segmentation using medical image processing chain optimization

Abstract

Surveyed literature shows many segmentation algorithms using different types of optimization methods. These methods were used to minimize or maximize objective functions of entropy, similarity, clustering, contour, or thresholding. These specially developed functions target specific feature or step in the presented segmentation algorithms. To the best of our knowledge, this thesis is the first research work that uses an optimizer to build and optimize parameters of a full sequence of image processing chain. This thesis presents a universal algorithm that uses three images and their corresponding gold images to train the framework. The optimization algorithm explores the search space for the best sequence of the image processing chain to segment the targeted feature. Experiments indicate that using differential evolution to build Image processing chain (IPC) out of forty-five algorithms increases the segmentation performance of basic thresholding algorithms ranging from 2% to 78%.