Performance Evaluation of the TINA Medical Image Segmentation Algorithm on Brainweb Simulated Images

This memo describes the performance evaluation of the TINA medical image segmentation algorithm described in Memo 2004-009 when applied to simulated images produced by the Brainweb MRI simulator. In order to allow Monte-Carlo experiments to be performed using independent image noise fields, and to avoid problems introduced by the presence of histogram artefacts in the Brainweb simulated images as identified in TINA Memo no 2008-002, images were re-simulated from the Brainweb tissue phantoms. Monte-Carlo experiments were performed by applying the segmentation algorithm both with and without the use of gradient terms. The results were evaluated by measuring the rates of voxel misclassification, treating the algorithm as a simple classifier, and by measuring the χ of the tissue volume estimates, in order to evaluate the accuracy of partial volume estimation. We conclude that the TINA medical image segmentation algorithm achieves misclassification rates close to the Bayes error for intensity-based segmentation, and that the use of gradient terms improves the accuracy of tissue volume estimates in partial volume voxels by an average of 31%.