Biomedical Engineering Reference
In-Depth Information
MLS calculation is only performed in a small region of the volume, rather than
throughout the whole volume, making the computational cost proportional to
the object surface area [36]. As opposed to many interpolation schemes, the
MLS method is stable with respect to noise and imperfect registrations [46]. Our
implementation also allows for small intensity attenuation artifacts between the
multiple scans thereby providing gain-correction. The distance-based weighting
employed in our method ensures that the contributions from each scan are prop-
erly merged into the final result. If a slice of data from one scan is closer to a
point of interest on the model, the information from this scan will contribute
more heavily to determining the location of the point.
To the best of our knowledge there is no previous work on creating de-
formable models directly from multiple volume datasets. While there has been
previous work on 3D level set segmentation and reconstruction [5, 6, 8, 41, 47], it
has not been based on multiple volume datasets. However, 3D models have been
generated from multiple range maps [29, 36, 48, 49], but the 2D nature of these
approaches is significantly different from the 3D problem being addressed here.
The most relevant related projects involve merging multiple volumes to produce
a single high-resolution volume dataset [50,51], and extracting edge information
from a single nonuniform volume [52]. Our work does not attempt to produce a
high-resolution merging of the input data. Instead, our contribution stands apart
from previous work because it deforms a model based on local edge information
derived from multiple nonuniform volume datasets.
We have demonstrated the effectiveness of our approach on three multi-
scan datasets. The first two examples are derived from a single high-resolution
volume dataset that has been subsampled in the
X
,
Y
, and
Z
directions. Since
these nonuniform scans are extracted from a single dataset, they are therefore
perfectly aligned. The first scan is derived from a high-resolution MR scan of a
12-day-old mouse embryo, which has already had its outer skin isolated with a
previous segmentation process. The second example is generated from a laser
scan reconstruction of a figurine. The third example consists of multiple MR
scans of a zucchini that have been imperfectly aligned by hand. The first two
examples show that our method is able to perform level set segmentation from
multiple nonuniform scans of an object, picking up and merging features only
found in one of the scans. The second example demonstrates that our method
generates satisfactory results, even when there are misalignments in the regis-
tration.
