A comparative performance evaluation of various approaches for liver segmentation from SPIR images

Developing a robust method for liver segmentation from magnetic resonance images is a challenging task because of the similar intensity values between adjacent organs, the geometrically complex liver structure, and injection of contrast media. Most importantly, a high anatomical variability of a healthy or diseased liver is a major challenge in defining the exact boundaries of the liver. Several artifacts of pulsation, motion, and partial volume effects are also among the variety of factors that make automatic liver segmentation difficult. In this paper, we present an overview of liver segmentation methods in magnetic resonance images and show comparative results of seven different pseudo-3D liver segmentation approaches chosen from deterministic (K-means-based), probabilistic (Gaussian model-based), supervised neural network (multilayer perceptron-based), and deformable model-based (level set) segmentation methods. The results of quantitative and qualitative analyses using sensitivity, specificity, and accuracy metrics show that the multilayer perceptron-based approach and a level set-based approach, both of which use distance regularization terms and signed pressure force function, are the most successful methods for liver segmentation from spectral presaturation inversion recovery (SPIR) images. However, the multilayer perceptron-based segmentation method has a higher computational cost. The automatic method using the distance regularized level set evolution with signed pressure force function avoids the sensitivity of a user-defined initial contour for each slice, gives the most efficient results for liver segmentation after the preprocessing steps, and also requires less computational time.