Towards Automatic Medical Image Segmentation using Shape Particle Filters

L. Fischer, R. Donner, F. Kainberger, G. Langs

Purpose: Segmentation approaches based on Shape Particle Filters deliver promising results for the localization of anatomical structures in medical images. They are used for the segmentation of human vertebrae, lungs and hearts, and are especially well suited to cope with the high levels of noise encountered in MR data and overlapping structures with ambiguous appearance in radiographs. They require a region template of the appearance features which allow to estimate the confidence in the hypotheses generated during the search. Currently these templates are created manually, which introduces a bias, and leads to particularly sub-optimal results in complex anatomy.

Methods and Materials: A Differential Evolution based segmentation scheme where the optimal number of template regions is derived automatically from a set of training images is employed. The method was evaluated using a) manual and b) automatic region maps on two annotated data sets: 1) hand radiographs, 2) heart MRI slices. The distance between segmentation results and ground truth was measured.

Results: The median/mean/std pixel distance for 1) was a) 10.12/12.51/9.88 and b) 4.94/7.21/6.92 and for 2) was a) 5.80/8.96/11.26 and b) 4.36/5.10/3.59. Results using an automatic region map outperformed those with manual region maps in all experiments.

Conclusion: By using automatically derived region maps, the laborious estimation of suitable manual region maps through trial and error is eliminated, paving the way to rapid application of Shape Particle Filters in clinical scenarios. Future work will focus optimal feature selection and extending the method to 3D.