A probabilistic model for classifying segmented images

In this work we introduce a probabilistic model for classifying segmented images. The proposed classifier is very general and it can deal both with images that were segmented with deterministic algorithms, such as the k-means algorithm, and with probabilistic clustering approaches, such as the Hidden Markov Random Field (HMRF) algorithm. Similarly, our model can be used on either binary images or on images that contain multiple clustering labels as well as on images with any cluster boundaries (sharp, fuzzy or irregular). We tested our classifier on real fMRI images and showed that it outperforms the region-based Maximum Likelihood k-means classifier. Furthermore, we showed that higher classification rates are obtained when the images are segmented using a probabilistic HMRF algorithm compared to deterministic k-means method.