Sample Sufficiency and Number of Modes to Retain in Statistical Shape Modelling

摘要: Statistical shape modelling is a popular technique in medical imaging, but the issue of sample size sufficiency not generally considered. Also number principal modes retained often chosen simply to cover percentage total variance. We show that these simple rules are unreliable. propose new method uses bootstrap replication and t-test comparison with noise decide whether each mode direction has stabilised. establish correspondence by minimising distance between space spanned replicates their mean. By retaining only stable modes, our distinguishes real anatomical variation from dominated random noise. This provides lower stopping rule when small converges as increases. use this convergence determine sufficiency. For validation we synthetic datasets left ventricle generated known structural added Our detected correct retain where other methods failed. The were also tested on 2D (22 points) 3D (500 face data, 24 70 being reached at approximately 50 150 samples respectively. database (527 points), 319 sufficient, level can around 55 modes. principled foundation for appropriate selection determination statistical modelling.