Subject-Specific prior shape knowledge in feature-oriented probability maps for fully automatized liver segmentation in MR volume data

• Published in: Pattern recognition Vol. 84; pp. 288 - 300
• Main Authors: Gloger, Oliver, Tönnies, Klaus
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.12.2018
• Subjects: 3D prior shape level set segmentation; Bayesian probability; Expectation maximization; Normalized cross correlation; Principal component analysis; Subject-specific shape model; Expectation maximization; 3D prior shape level set segmentation; Subject-specific shape model; Bayesian probability; Normalized cross correlation; Principal component analysis
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2018.07.018

•Fully automatized liver segmentation approach for native MR volume data•Subject-specific prior liver shape incorporation into level set segmentation•Liver tissue-specific probability map generation combining all MR contrasts•Consideration of inner-organ MR-differences and recognition of fat livers and cysts•Novel alignment and attraction techniques for exact liver delineation Liver segmentation and volumetry in native MR-volume data is an important topic in epidemiological research. Manual liver segmentation is extremely time-consuming and often infeasible requiring automatized methods. Automatic liver segmentation is challenging because of the large variability in liver shape and appearance and the low contrast to neighboring organs. We present a fully automatized liver segmentation framework that uses a sequence of modules based on individualized model knowledge on liver appearance and shape. Liver probability maps are computed that incorporate organ-specific features like MR-intensity distributions, inner-organ MR-differences and liver positions. Probability map generation differentiates automatically between fatty and non-fatty livers. Moreover, we improve an existing technique for prior shape level set segmentation to delineate the liver in tissue-specific liver probability maps and to recognize cystic hepatic tissue. Dice coefficients of 0.937 and low volumetric errors on 35 test data sets confirm the robust segmentation quality of the framework. [Display omitted]