Automatic Annotation of Radiographs using Random Forest Regression Voting for Building Statistical Models for Skeletal Maturity

Statistical Models of Shape and Appearance require annotation of the bones of the hand of children and young adults. Due to very large variation in the shape and appearance of these bones, automatic annotation is particularly challenging. Statistical Models of Shape and Appearance have been found useful in several medical image analysis and other applications. In this work we locate sparse points on the bones of the hand with an automatic system which uses a Constrained Local Model with Random Forest Regression Voting. These sparse points were then used as input to a groupwise registration algorithm. The control point of the groupwise algorithm can then be used to propagate
manually annotated points to other images. The resulting propagation may be used to build Statistical models. By analysing performance on dataset of 537 digitized images of normal children we achieved an automatic annotation
accuracy of a mean point to curve error of 0.94mm ± 0.01 and a median error 0.92mm.