Non-deterministic optimization using differential evolution algorithm to launch seeds for liver segmentation in MDCT

Abstract

The automatic segmentation of liver in Multi Detector Computed Tomography (MDCT) is considered a complex task due to the similarity and proximity of the liver with other soft tissues. Although some techniques already achieve good results, their deterministic characteristic might prevent the feasibility of the algorithm in complex situations. Thus, this paper purpose the use of Differential Evolution (DE) as a nondeterministic evolutionary algorithm in order to optimize the launching of seeds for region growing segmentation algorithms. In order to achieve this goal, a fitness function, based on intensity, contrast, morphology and localization features of the liver, along with a full exam analysis methodology were developed. Ten MDCT exams were then submitted to the method, resulting in 70% of exams with at least 85% of correctly seeded liver. However, some exams presented high false-positive seeding due to the complexity of the images. In conclusion, the purposed use of DE has shown promising results on launching seeds inside the liver on MDCT images.