Pranav Bhagirath

48 Chapter 3 Table 3. A brief summary of algorithms that were evaluated on the proposed framework. Algorithm Technique Strengths and weaknesses Key features Interaction AIT : Lara et al. Otsu, support vector machines and level-sets Post-processing improves results but increases running time Otsu with two tissue classes. User selects seed in blood-pool Semi- automatic UPF : Albà et al. Region-growing and morphology Shapes uncharacteristic of scar are deleted but requires initialization for every slice Two seeds, for healthy and scar, per slice. Region-labeling step ensures smoothness, filling gaps Semi- automatic MCG : Karimaghaloo et al. Conditional random fields Hierarchical approach with two levels of processing, but uses statistics on a small neighborhood Posterior distribution model estimated with a direct map and not Gaussian during training Automatic MV : Hennemuth et al. EM-algorithm and Watershed Transformation No fixed intensity model and the best-fit model is selected, but over-fitting can be an issue Automated seed-selection in watershed process. Gaussian-mixture or Rician–Gaussian models for fitting intensities with EM algorithm Semi- automatic KCL : Karim et al. Graph-cuts with EM- algorithm Computes a globally optimal segmentation, but can sometimes reject good candidates. Gaussian-mixture model fits intensities with EM algorithm using three tissue classes Semi- automatic n - SD n standard deviations from healthy tissue (n = 2, 3, 4, 5, 6) Simple to implement, but baseline is subjective Only involves thresholding, no region-growing as FWHM Semi- automatic FWHM 50% of user-selected hyper-enhanced myocardium Validated with histology in literature but was first used to describe a phenomenon in signal analysis Computed threshold used for region-growing from user-selected seed locations in each slice Semi- automatic Institution abbreviations: AIT - Alma IT Systems, and - Universitat Pompeu Fabra, MCG - McGill University, MV - Mevis Fraunhofer, KCL - King’s College London.