Information Technology Reference
In-Depth Information
and passed through a detection
filter. The candidates which successfully pass
through the detection
cation.
The watershed algorithm is again applied to detect the metastasis candidates.
Watershed algorithm views the gradient of the image intensity as a topographic
surface in order to extract relatively homogeneous regions of the image called
catchment basins, some of which will be candidates for lesions. The algorithm can
be adapted for both lytic and sclerotic lesions. For lytic lesions, low intensity
regions surrounded by high intensity regions are detected. Similarly, for sclerotic
lesions, high intensity regions surrounded by low intensity regions are detected.
Example results of the watershed algorithm are shown in Fig.
6
.
We then address the over-segmentation problem in watershed with a post-
watershed merging routine using a graph-cuts strategy [
43
]. Without loss of gen-
erality, we use the sclerotic lesion detection to describe the graph-cut strategy. We
filter are then sent to the next stage for classi
first initialize each watershed region with a foreground (F) or background (B) label.
There are two types of foreground regions: those in the cortical bone region and
those in the medullary regions. Any region that has intensity 100 HU higher than its
surrounding regions (cortical or medullary) will be initialized as F. The rest of the
regions are initialized as B. The regions and their neighbors are fed into a graph-cuts
merging routine.
An adjacency graph for watershed regions is constructed by representing
adjacent regions as nodes connected by edges [
44
]. The technique partitions the set
of nodes into two disjoint sets F and B in a manner that minimizes an energy
function,
Fig. 6 Candidate detection and segmentation. a CT image of a vertebra with sclerotic lesions;
b watershed result; c graph cut result after watershed; d candidate sclerotic lesions; e 3D
segmented sclerotic lesions; f CT image of a vertebra with lytic lesions; g watershed result;
h candidate lytic lesions; and i 3D segmented lytic lesions
