Superpixel and entropy-based multi-atlas fusion framework for the segmentation of x-ray images

Abstract

X-ray images segmentation can be useful to aid in accurate diagnosis or faithful 3D bone reconstruction but remains a challenging and complex task, particularly when dealing with large and complex anatomical structures such as the human pelvic bone. In this paper, we propose a multi-atlas fusion framework to automatically segment the human pelvic structure from 45 or 135-degree oblique X-ray radiographic images. Unlike most atlas-based approach, this method combines a data set of a priori segmented X-ray images of the human pelvis (or multiatlas) to generate an adaptive superpixel map in order to take efficiently into account both the imaging pose variability along with the interpatient (bone) shape non-linear variability. In addition, we propose a new label propagation or fusion step based on the variation of information criterion for integrating the multi-atlas information into the final consensus segmentation. We thoroughly evaluated the method on 30 manually segmented 45 or 135 degree oblique X-ray radiographic images data set by performing a leave-one-out study. Compared to the manual gold standard segmentations, the accuracy of our automatic segmentation approach is 85% which remains in the error range of manual segmentations due to the inter intra/observer variability.