Structure-preserving composite model for skin lesion segmentation
A structure-preserving composite model for skin lesion segmentation includes partitioning a dermoscopic image into superpixels at a first scale. Each superpixel is a vertex on a graph defined by color coordinates and spatial coordinates, and represents a number of pixels of the dermoscopic image according to the first scale. Further, constructing a plurality of k background templates by k-means clustering selected ones of the superpixels in space and color. Additionally, generating sparse representations of the plurality of superpixels based on the plurality of background templates. Also, calculating a reconstruction error for each superpixel by comparison of its sparse representation to its original color coordinates and spatial coordinates. Furthermore, outputting a confidence map that identifies each pixel of the dermoscopic image as belonging or not belonging to a skin lesion, based on the reconstruction errors of the representative superpixels.
1. An apparatus comprising:
means for outputting a confidence map that identifies each pixel of a dermoscopic image as belonging or not belonging to a skin lesion,
wherein the means comprises:
a first software module for partitioning the dermoscopic image into a plurality of superpixels,
a second software module for constructing a plurality of background templates by clustering similar superpixels from at least one corner superpixel of the dermoscopic image,
a third software module for generating sparse representations of the plurality of superpixels based on the plurality of background templates,
a fourth software module for calculating reconstruction errors of the sparse representations by comparison of each sparsely represented superpixel to color coordinates and spatial coordinates of a corresponding original superpixel, and
a fifth software module for segmenting the plurality of superpixels as belonging or not belonging to the skin lesion, based on the respective reconstruction error of each of the plurality of superpixels.
2. The apparatus of claim 1 wherein the first software module partitions the dermoscopic image by implementing Simple Linear Iterative Clustering.
3. The apparatus of claim 1 wherein the second software module constructs the plurality of background templates by k-means clustering of the selected ones of the plurality of superpixels.
4. The apparatus of claim 3 wherein k=3.
5. The apparatus of claim 1 wherein the third software module generates sparse representations by solving the optimization problem
argmin
x
Y
-
BX
F
2
+
λ
1
2
Tr
(
XLX
T
)
+
λ
2
X
2
,
1
wherein Y is the matrix of superpixels, B is the matrix of background templates, X is the matrix of sparse representations, L is the Laplacian of a graph of the superpixels, λ 1 and λ 2 are pre-determined regularization parameters, and Tr( ) indicates a trace function.
6. The apparatus of claim 1 wherein the fourth software module calculates the reconstruction errors ε i s according to ε i s =∥y i s −B s x i s ∥.
7. The apparatus of claim 1 wherein the fifth software module segments each of the plurality of superpixels as belonging to the skin lesion in case a saliency value of that superpixel exceeds a mean saliency value of the plurality of superpixels, or as not belonging to the skin lesion in case the saliency value of that superpixel does not exceed the mean saliency value of the plurality of superpixels.
8. A computer program product comprising a computer readable storage medium embodying computer executable instructions which when executed by a computer cause the computer to facilitate the method of:
partitioning a dermoscopic image into a plurality of superpixels at a first scale, wherein each superpixel is a vertex on a graph defined by color coordinates and spatial coordinates, each of the plurality of superpixels representing a number of pixels of the dermoscopic image according to the first scale;
constructing a plurality of k background templates by k-means clustering selected ones of the superpixels in space and color from at least one corner superpixel of the dermoscopic image;
generating sparse representations of the plurality of superpixels based on the plurality of background templates;
calculating a reconstruction error for each superpixel by comparison of its sparse representation to its original color coordinates and spatial coordinates; and
outputting a confidence map that identifies each pixel of the dermoscopic image as belonging or not belonging to a skin lesion, based on the reconstruction errors of the representative superpixels.