Object localization in medical images based on graphical model with contrast and interest-region terms

In this paper, we propose a novel method for object localization, generally applicable to medical images in which the objects can be distinguished from the background mainly based on feature differences. We design a new CRF model with additional contrast and interest-region potentials, which encode the higher-order contextual information between regions, on the global and structural levels. We also propose a sparse-coding based classification approach for the interest-region detection with discriminative dictionaries, to serve as a second opinion for more accurate region labeling. We evaluate our object localization method on two medical imaging applications: lesion dissimilarity on thoracic PET-CT images, and cell segmentation on microscopic images. Our evaluations show higher performance when comparing to recently reported approaches.