Level sets-based image segmentation approach using statistical shape priors

• Published in: Applied mathematics and computation Vol. 340; pp. 164 - 179
• Main Authors: Eltanboly, Ahmed, Ghazal, Mohammed, Hajjdiab, Hassan, Shalaby, Ahmed, Switala, Andy, Mahmoud, Ali, Sahoo, Prasanna, El-Azab, Magdi, El-Baz, Ayman
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01-01-2019
• Subjects: Expectation-Maximization (EM); Linear combination of Gaussians (LCG); Probability density function (PDF); Shape prior; Expectation-Maximization (EM); Probability density function (PDF); Linear combination of Gaussians (LCG); Shape prior
• ISSN: 0096-3003; 1873-5649
• DOI: 10.1016/j.amc.2018.05.064

A robust 3-D segmentation technique incorporated with the level sets concept and based on both shape and intensity constraints is introduced. A partial differential equation (PDE) is derived to describe the evolution of the level set contours. This PDE does not contain weighting parameters that need to be tuned, which overcomes the drawbacks of other PDE approaches. The shape information is collected from a set of co-aligned manually segmented contours of the training data. A promising statistical approach is used to get the distribution of the intensity gray values. The introduced statistical approach is built by modeling the empirical PDF (normalized histogram of occurrences) for the intensity level distribution with a linear combination of Gaussians (LCG) incorporating both negative and positive components. An Expectation-Maximization (EM) algorithm is modified to deal with the LCGs, and we also proposed an EM-based sequential technique to acquire a close initial LCG approximation for the modified EM algorithm to start with. The PDF of the intensity levels is incorporated in the speed function of the moving level set to specify the evolution direction. Experimental results show how accurately the approach is in segmenting various types of 2-D and 3-D datasets comprising medical images.