Hierarchical Multi-Atlas Segmentation Using Label-Specific Embeddings, Target-Specific Templates and Patch Refinement

• Published in: Patch-Based Techniques in Medical Imaging Vol. 9993; pp. 84 - 91
• Main Authors: Arthofer, Christoph, Morgan, Paul S., Pitiot, Alain
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 01.01.2016; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Atlas Image; Deformation Field; Graphics programming; Image processing; Manifold Space; Pattern recognition; Target Image; Trilinear Interpolation
• ISBN: 9783319471174; 3319471171
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-47118-1_11

Patch-based Multi-Atlas Segmentation methods typically transform a priori expert delineations of atlas images onto a new target image where they are fused based on local patch similarities. To improve efficiency and accuracy, we build a population template offline and a Target-Specific Template (TST) at runtime, which act as intermediate steps. At the regional level, we build a manifold for each label rather than globally, and from the deformation fields rather than from images, in order to better model the geometry of atlases and target with the aim to improve the similarity between TST and target. At the local level, we further refine the resulting weights in those areas prone to registration errors with patches sampled from the warped- and target-gray matter probability images. We evaluated our approach on the standard NIREP dataset for which it achieved state-of-the-art performance while being up to 16 times faster than competing approaches.