Stability Testing of the TINA Medical Image Segmentation Algorithm

This memo describes the stability testing of the TINA medical image segmentation algorithm de-scribed in Memo 2004-009. The performance of the algorithm when applied to a common medical image segmentation problem, the segmentation of structural MR image volumes of the brain into CSF, grey matter and white matter regions, was evaluated. We examine the stability of the optimised tissue model after the optimisation is initiated from multiple random starting points. This method allows comparative evaluation of the results without the need to define a gold-standard segmentation. A straw-man algorithm, based on simplex optimisation of a χ2 histogram fit of the model to the images, is used to provide a baseline for this comparison. The superiority of the EM-based algorithm over this simpler alternative, and the improvements gained through the inclusion of gradient information in the tissue model, are demonstrated.