Figure: (a, b) Two MR scans of a person's
knee. Both images have high resolution in-plane, but have about one
quarter the resolution between planes. (c) A single SurfaceNet is
built around an object and then relaxed, producing a smooth surface,
free of terracing. (d) Two nets are built from two orthogonal scans
and relaxed.
The generation of three-dimensional models of anatomical structures from medical imagery is important for applications such as surgical simulation, planning, and image-guided surgery. An internal scan typically consists of high-resolution data in the imaging plane and significantly lower resolution between imaging slices. The lack of high-resolution information along the scanning direction causes aliasing or terracing artifacts in anatomical surface models, which can be distracting or misleading to surgeons. For surgical simulation, the terraces subtract from the realism of the visualization and create very noticeable ridges when using haptics to feel the object's surface. These terracing artifacts can be reduced by increasing the resolution of the scan. However, for CT scans, higher resolution between imaging planes subjects patients to a higher dose of radiation. For MR scans, longer scan times are necessary to achieve higher resolution, which is more costly and is more difficult for the patient, who must remain absolutely still during image acquisition.
For clinical practice, scans are usually acquired in more than one orthogonal direction. For example, instead of acquiring a single very high resolution sagittal MR scan, lower resolution sagittal and axial scans may be acquired (see Fig. 1a,b). Surgeons and radiologists use information from both acquisitions for diagnosis, surgical guidance, and treatment. Similarly, we are interested in combining the information from two scans to produce three dimensional models of internal structures that have higher resolution than models created from either of the scans alone. The method proposed here is an extension of the Constrained Elastic SurfaceNet described in [2], which generates models from a single scan.