Analysis of parameters' effects in semi-automated knee cartilage segmentation model: Data from the osteoarthritis initiative

Abstract

Unlike automated segmentation, the accuracy of semi-automated segmentation is affected by pertinent parameters such as observer, type of methods and type of cartilage. In this paper, we investigated the effect of these parameters on segmentation results. Based on Dice similarity index obtained from fifteen normal and ten diseased magnetic resonance images, a parameter estimation model was constructed to study the impact of each parameter. Then, we conducted deviance test to verify the effect's significance. Our result showed that implementation of the proposed segmentation model would introduce positive effect (+0.12) on reproducibility compared to conventional random walks model. Furthermore, we have found intriguing results indicating cartilage normality has diminished effect on reproducibility and tibial cartilage's result could be influenced by external factors as well. Lastly, our findings highlighted on the necessity of refinement for semi-automated segmentation.