Assessment of MRF based joint scale selection and segmentation for 3D liver vessel segmentation task

Vessel segmentation plays an important role in medical image analysis. Irrespective of the modality used, the common challenge in all vessel tracking methods is scale variability, in other words, the dependence on the size of the vessels, which is unknown a priori. Despite few approaches that attempts to perform scale selection and segmentation simultaneously, the common approach is to perform multiscale analysis and fuse the results afterwards via a scale selection mechanism. Recently, Mirzaalian et al. proposed to use MRFs for joint scale selection and vessel segmentation in 2D retinal images. In this study, we have assessed the 3D version of this method in comparison with the conventional multiscale approach augmented with novel automatic threshold selection and image guided morphological filtering, using the well-known Hessian based method. The assessment has been done quantitatively using IRCAD dataset and qualitatively by studying the output vessel masks. The results indicate that the MRF based approach does not improve the results significantly in 3D liver vessel segmentation task compared to conventional multiscale approach.