Medical imaging is being increasingly used, and the demand for automatic segmentation of structures of interest grow.

In this work a model-based segmentation method is implemented: An active appearance model based on principal component analysis with a level-set representation of shape is utilized in an iterative algorithm for segmentation of 3-D images. The automatic segmentation algorithm incorporates prior knowledge to predict how to correct model and pose parameters in order to achieve a better fit of the model to the target image.

The segmentation method is tested on 42 prostate MR images and 27 CT images of the L4 vertebra in a leave-one-out cross-validation framework. The automatic segmentations are compared to manual reference segmentations. A median Dice kappa of 0.81 is achieved for both structures.

The algorithm performs similar to previously described methods, but in some cases it fails to determine the correct size of the prostate. Also the appearance model is not large enough to fully segment the vertebral processes. The algorithm is sensitive to the initial location of the average model in the target image. The active appearance model presented can be applied on any imaging modality and any structure of interest if the shape of the structure is not too variable.