Which Metrics Should Be Used in Non-linear Registration Evaluation?
Non-linear registration is an essential step in neuroimaging, influencing both structural and functional analyses. Although important, how different registration methods influence the results of these analyses is poorly known, with the metrics used to com
- PDF / 681,190 Bytes
- 8 Pages / 439.363 x 666.131 pts Page_size
- 88 Downloads / 215 Views
ion
Analysis of medical data often requires a precise alignment of subjects’ scans with a common space or atlas. This alignment allows the comparison of data across time, subject, image type, and condition, while also allowing its segmentation into different anatomical regions, or find meaningful patterns between different groups [7]. Due to high inter-subject variability, a non-linear deformation of a subject’s scan is typically applied to best conform this image to a reference standard. This type of deformation allows for complex modulation, such as elastic or fluid deformations. The quantification of the accuracy of a non-linear registration is intrinsically complex. The main reason being the lack of a ground truth for the validation of different methods [8]. While in a rigid registration only three non-colinear landmarks are required, in non-linear registration a dense mesh of landmarks is needed [12]. To provide ground truth data the EMPIRE 10 challenge [9] provided correspondences of 100 annotated landmark pairs to distinguish between registration algorithms. However, the precision of the evaluation is still limited by the number of correspondence points. Furthermore, this study focused on intra-subject thoracic CT which may not fully apply across inter-subject studies or other modalities and structures. Two other main approaches exist to evaluate non-linear registration methods: one based on the simulation of deformation fields; and the other in the evaluation of manually segmented regions in different subjects. c Springer International Publishing Switzerland 2015 N. Navab et al. (Eds.): MICCAI 2015, Part II, LNCS 9350, pp. 388–395, 2015. DOI: 10.1007/978-3-319-24571-3_47
Metrics for Registration Evaluation
389
In the former, several types of deformation field simulations are suggested, based in the deformation of control points, biomechanical models [2],[3],[13] or Jacobian maps [6],[11]. These techniques can struggle in that they are not capable of simulating the totality of inter-subject variation, nor the artefacts present in medical images. In the latter, a database of different subjects with both an anatomical image and manually segmented volumes of interest (VOI) is typically used, in which each subject is registered to a randomly chosen subject, or average image, and the calculated deformation applied to the VOI map [4],[15]. In this way, the registered VOI map of the source subject can be compared with the VOIs of the target subject. Typically, to evaluate the different methods a labelling metric is used: Dice coefficients [7]; volume overlap [7],[14]; Jaccard index [10],[12]. This analysis assumes that the metrics evolve in the same way as the unknown deformation field. To our knowledge no study has been performed that effectively compare such metrics with a ground truth such as the true deformation field. In this study we simulate pairs of deformation fields target images, and evaluate the relation between the different similarity metrics and the true deformation field. We explore what proportion of the varianc
Data Loading...
