Specificity, Generalisation and Registration
Date: Thursday, August 18 @ 04:43:51 BST
Topic: Technical Notes
e currently work on heavy and extensive experiments where registration gets evaluated using models of appearance. Models of the brains are built from progressively eroded registration and from these models we derive Specificity and Generalisability (Generalisation ability) values. The figures below show that our method successfully discerns good registration from a worse one within a wide range of deformations (perturbations) of the correct solution. Lower values indicate a better registration. 4 types of shuffle distances (including the Euclidean distance) are used in the process.
It is worth pointing out that our method requires no ground truth to be provided for evaluation and it was shown to be tightly-correlated to overlap-based measures, which require manual segmentation of data labels.
See related paper:
Carole Twining, Tim Cootes, Stephen Marsland, Vladimir Petrovic, Roy Schestowitz, and Chris Taylor. A Unified Information-Theoretic Approach to Groupwise Non-Rigid Registration and Model Building. Presented in Information Processing in Medical Images (IPMI), Lecture Notes in Computer Science, vol. 3565, pp. (PDF, BibTeX)