In this work we introduce two different techniques for the global optimization of surfaces and apply them to the task of finding the optimal stitching seam between neighboring and overlapping 3D ultrasound volumes. Existing techniques for US mosaicing, based on interpolation or planar seams, introduce artifacts into the composite volume especially when using a large number of clinical scans. Our first method models the seam as a B-spline surface and treats its calculation as a shape optimization problem. In this case the optimal location of the surface-defining control points is a large scale constrained optimization problem, which is solved using a cooperatively coevolving particle swarm based approach. The second method treats the seam selection as a voxel labeling problem, where each voxel in the composite volume is labeled with its respective source volume. Therefore if we have N volumes, each voxel in the composite volume may be assigned one of the N labels. The optimal labeling, which implicitly defines a seam, minimizes the intensity and gradient difference between adjacent volumes The formulation is optimized using graphcuts, which guarantees that a global minimum is achieved due to the submodularity of the energy function. The final composite volume is constructed voxel-wise by taking the value of the source volume, which is designated by its label. Our application of this procedure is the construction of composite ultrasound image volumes for incorporation into an ultrasound simulator. These methods are validated on clinical US data acquired from obstetrics patients.

Date Published: 13 March 2013
PDF: 12 pages
Proc. SPIE 8669, Medical Imaging 2013: Image Processing, 866949 (13 March 2013); doi: 10.1117/12.2006997

Published in SPIE Proceedings Vol. 8669:
Medical Imaging 2013: Image Processing
Sebastien Ourselin; David R. Haynor, Editor(s)