| 主权项 |
1. A method for unsupervised segmentation of microscopic color image of unstained specimen and digital staining of segmented histological structures by using empirical kernel map-based nonlinear mapping of recorded microscopic image of unstained specimen onto reproducible kernel Hilbert space, factorization of mapped image constrained by nonnegativity and l0-norm of the binary {0, 1} sources (histological structures) and digital staining of factorized (segmented) histological structures comprising the following steps:
recording and storing microscopic color image of unstained specimen X, where X is nonnegative data matrix comprised of N=3 rows that correspond to gray scale images recorded at particular wavelengths corresponding to red, green and blue colors and T columns that correspond to observations at different spatial (pixel) locations, scaling the image data matrix by maximal element of X, xmax:
X=X/xmax [I]representing image data matrix X by linear mixture model:
X=AS [II] where AεR0+3×M stands for nonnegative mixture matrix comprised of M column vectors {am}m=1M that stand for spectral profiles of M histological structures present in the image X; S stands for M×T binary source matrix comprised of {0, 1} values such that element {smtε{0,1}}m,t=1M,T indicates presence (1) or absence (0) of the histological structure m at pixel location t.
using empirical kernel map for nonlinear mapping of X in [II] onto reproducible kernel Hilbert space Ψ(X)εR0+D×T:Ψ(X)=[κ(x1,v1)…κ(xT,v1)………κ(x1,vD)…κ(xT,vD)][III] where κ(xt,vd), t=1, . . . , T and d=1, . . . , D stands for positive symmetric kernel function and vd, d=1, D stand for basis vectors that approximately span the same space as pixels vectors: xt, t=1, . . . , T.
representing mapped matrix Ψ(X) by linear mixture model [IV]:
Ψ(X)=BS [IV] such that S is the same binary source matrix as in [II], while BεR0+D×M is mixing matrix in mapped space such that column vectors {bm}m=1M are mutually significantly less correlated than column vectors {am}m=1M in [II]. That enables discrimination of spectrally similar histological structures present in the image X.
applying sparseness and nonnegativity constrained matrix factorization (sNMF) algorithm to [IV], whereas sparseness constraint is based on indicator function of S such as lo quasi-norm of S, to obtain estimates of the presence/absence of histological structures {sm}m=1M:
{ŝm}m=1M=sNMF(Ψ(X)) [V] where, as in [II], M denotes number of histological structures present in the image X;
displaying segmented histological structures {ŝm}m=1M as black and white maps;digitally staining (coloring) segmented histological structures {ŝm}m=1M with predefined colors according to:
Y=CŜ [VI] where {cm}m=1M stand for predefined color vectors in RGB-color space.
displaying segmented histological structures as synthetic color (RGB) image Y. |
