Active contours and information theory for supervised segmentation on scalar images

In this paper, we present two models for supervised scalar image segmentation based on the active contours and information theory. First we propose to carry out a region competition by optimizing an energy designed to be minimal when the entropy of the inside and outside regions of the evolving active contour are close to those of a reference image. The probability density functions (pdfs) used by this model can be computed in a preprocessing step on a reference image. This substantially reduces the computational complexity making this model fast. On the other hand, this implies that the reference image and the image to segment have similar pdfs. When the pdfs are too different or both images are not from the same modality we propose a second segmentation model computationally more expensive but more robust to intensity differences. This second model is based on an information measure extensively used for image registration, the joint entropy. The performance of both models is demonstrated on a variety of 2D synthetic data and medical images. They are also compared in term of segmentation accuracy and computational cost with an entropy-based unsupervised segmentation model recently proposed.