Learning Correspondences in Knee MR Images from the Osteoarthritis Initiative

• Published in: Machine Learning in Medical Imaging Vol. 7588; pp. 218 - 225
• Main Authors: Guerrero, Ricardo, Donoghue, Claire R., Pizarro, Luis, Rueckert, Daniel
• Format: Book Chapter
• Language: English
• Published: Germany Springer Berlin / Heidelberg 2012; Springer Berlin Heidelberg
• Series: Lecture Notes in Computer Science
• Subjects: Artificial intelligence; Feature Correspondence; Image processing; Pattern recognition; Random Sample Consensus; Statistical Shape Model; Structure Tensor; Transformation Model
• ISBN: 9783642354274; 3642354270
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-642-35428-1_27

Registration is a powerful tool that allows mapping images in a common space in order to aid in their analysis. Accurate registration of images of the knee is challenging to achieve using intensity based registration algorithms. Problems arise due to large anatomical inter-subject differences which causes registrations to fail to converge to an accurate solution. In this work we propose learning correspondences in pairs of images to match self-similarity features, that describe images in terms of their local structure rather than their intensity. We use RANSAC as a robust model estimator. We show a substantial improvement in terms of mean error and standard deviation of 2.13mm and 2.47mm over intensity based registration methods, when comparing landmark alignment error.