Outlier Rejection for Adaptive Neonatal Segmentation

M. Jorge Cardoso¹, Andrew Melbourne¹, Giles S. Kendall², Nicola J. Robertson², Neil Marlow², Sebastien Ourselin¹

Volume estimation through automated segmentation can help predict neurodevelopmental outcome in babies born prematurely. However, automated segmentation techniques are hampered by lack of contrast, white matter (WM) abnormalities and anatomical variability. We propose an Expectation Maximisation (EM) segmentation algorithm with a prior over intensities and tissue proportions, a B0 inhomogeneity correction, a spatial homogeneity term and a robust outlier rejection technique that ignores unexpected intensity clusters. This technique significantly improves both the accuracy and the robustness of the segmentation to the presence of WM abnormalities and pathological variability when compared to a state-of-the-art EM-segmentation.