Active contours and information theoryfor supervised segmentation on scalar images

In this paper, we present two models for supervised scalar imagesegmentation based on the active contours and information theory.First we propose to carry out a region competition by optimizingan energy designed to be minimal when the entropy of the insideand outside regions of the evolving active contour are close to thoseof a reference image. The probability density functions (pdfs) usedby this model can be computed in a preprocessing step on a referenceimage. This substantially reduces the computational complexitymaking this model fast. On the other hand, this implies thatthe reference image and the image to segment have similar pdfs.When the pdfs are too different or both images are not from thesame modality we propose a second segmentation model computationallymore expensive but more robust to intensity differences.This second model is based on an information measure extensivelyused for image registration, the joint entropy. The performance ofboth models is demonstrated on a variety of 2D synthetic data andmedical images. They are also compared in term of segmentationaccuracy and computational cost with an entropy-based unsupervisedsegmentation model recently proposed.