Automatic and Semi-automatic Kidney segmentation has applications in the field of discovering disease and in planning for surgical resection [1]. Unfortunately the kidney and its neighbors are soft tissue and present the following challenges:
Soft tissue is deformable, and may not be the anticipated shape.
Most soft tissue does not contrast strongly against its background
Organs often have similar consistency and may be pushed against each other.
We reviewed both a 3D region growing and a non-uniform rational B-spline (NURBS) approach in our initial experiments, but later focused our efforts on extending and refining our region growing results. The NURBS approach involves beginning with a 3D spline model of an "average" kidney which a user initially orientates on the test kidney. It then iterates: adjusting its spline coefficients until the shape of the test kidney is matched. The matched NURBS was then checked against a hand-segmented copy of the test kidney, and the error was recorded. Since the approach presumes the shape of a kidney to start, it matched healthy kidneys well, and would yielded "kidney shaped" results. The major drawback to this approach was that it had significant difficulty shaping itself to include tumors and other renal irregularities.
Consequently, we focused on a region growing approach: here tumors can be "grown into" and detected. The associated downside is that the algorithm can also grow into adjacent organs, and falsely report the kidney shape as a result. Instead of positioning an entire model kidney into place, this approach only required a human to designate a seed point. From this point, the region is grown from a 3x3x3 voxel volume wherein the median of the 8-connected intra-slice pixels and the 2-connected inter-slice pixels is measured. If the selected pixel is within a threshold of the calculated median value, then the voxel is added to the volume.
|
|
|