Bone texture characterization with fisher encoding of local descriptors

Bone texture characterization is important for differentiating osteoporotic and healthy subjects. Automated classification is however very challenging due to the high degree of visual similarity between the two types of images. In this paper, we propose to describe the bone textures by extracting dense sets of local descriptors and encoding them with the improved Fisher vector (IFV). Compared to the standard bag-of-visual-words (BoW) model, Fisher encoding is more discriminative by representing the distribution of local descriptors in addition to the occurrence frequencies. Our method is evaluated on the ISBI 2014 challenge dataset of bone texture characterization, and we demonstrate excellent classification performance compared to the challenge entries and large improvement over the BoW model.