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 multi-atlas) to generate an adaptive superpixel map in order to take efficiently into account both the imaging pose variability along with the inter-patient (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.