User-Friendly Simultaneous Tomographic Reconstruction and Segmentation with Class Priors

Joint CT Reconstruction and Segmentation With Discriminative Dictionary Learning

“…The proposed method provides both a reconstruction and segmentation of the object, and so should the methods we want to compare it against. We choose one of the latest SRS methods, from [8], as method 1; most other techniques provide only a solution to one of the two unknowns. While we, in principle, could combine separate reconstruction and segmentation methods, many of them are not suited for textured images which would result in an unfair comparison.…”

“…The segmentation is then obtained by assigning the kth label with the most evidence, similar to (6). Method 1: This method is from [8] and it represents a non-dictionary based approach. It is a simplified version of the algorithm from [6], and the required prior information is the mean attenuation coefficient for each of the classes.…”

“…The underlying assumption is that the relationship between the reconstructed intensities and the class probabilities can be modelled using Gaussian distributions with joint variance. See [8] for details.…”

“…, using the notation from (8). Initial Class Probability: We first estimate a sparse representation of the initial reconstruction using the dictionary D:…”

“…Traditionally, image reconstruction and segmentation are performed as two separate steps -even though reconstruction errors may propagate into misclassifications. More recent techniques perform the two steps simultaneously through a joint problem formulation, called Simultaneous Reconstruction and Segmentation (SRS), see, e.g., [4], [5], [6], [7], [8]. These methods have a potential advantage because the segmentation acts as a regularizer of the inverse problem; a potential disadvantage is that they may need more algorithm parameters.…”