On the use of boundary gradient for the analysis of MR wrist bones volumes segmentation

Image segmentation partitions a digital image into regions of interest, by taking into account the information content and local properties of homogeneity and topological connectivity. The evaluation of segmentation quality is usually based on supervised ground-truth images, which are provided by experts and used for comparison with segmented images. Indeed, intra-rater and inter-rater reliability affect the accuracy of such reference information. In addition to subjectivity, the skill of the experts and their knowledge could lead to bias. To achieve an objective performance measure, an automatic method for unsupervised evaluation of volume segmentation is proposed, which aims at finding a repetitive and user-independent evaluation of segmented results. The proposed approach leverages on the integration of volume- and boundary-based analysis. The quality of the segmentation is estimated on the basis of gradient properties, measured in the volume and focused on the voxels of the border. The segmentation is evaluated through a simple fuzzy index based on the resulting gradient. The experimental phase is conducted on a set of Magnetic Resonance volumes of the wrist district.