Product

Spectral Unmixing for Hyperspectral Images of Breast Cancer Preparations

September 27, 2010

Breast carcinoma is the second most common form of cancer among women [1]. Multicolour fluorescent in-situ hybridisation (M-FISH) is a common method for staging breast carcinoma. The interpretation of M-FISH images is complicated by two effects: (i) the emission spectra of the fluorochrome marked DNA probes overlap each other, and (ii) healthy tissues auto fluorescence. In this paper spectral unmixing is applied to hyper-spectral images of M-FISH preparations to provide false colour images with higher contrast and better visual separability than standard RGB images. We implement spectral unmixing (SU) by combining orthogonal projection analysis.

(OPA), alternating Least Squares (ALS), simple-to-use interactive self-modelling mixture analysis (SIMPLISMA) and VARIMAX. SU is applied to the data to unmix the spectra into the desired fluorochrome signals and unwanted components, thus enabling the reduction of tissue autofluorescence and resolving spectral overlap in the emission spectra. The results show that (i) spectral unmixing reduces the intensity caused by tissue autofluorescence by up to 78% and (ii) the image contrast is enhanced by reducing the overlap between the emission spectra.

Spectral imaging

Spectral imaging (SI) acquires spatially resolved images of a measurement sample at different wavelengths and combines them into a three dimensional image cube. The two main approaches for the acquisition of hyperspectral image data are wavelength scanning and spatial scanning. Wavelength scanning methods take images at a certain wavelength range and both spatial axes at once. The spectral information is acquired sequentially. Spatial scanning techniques use imaging spectrographs, e.g. prism-grating-prism combinations that disperse incident light of a single line of an object into its spectral components and project it onto a two dimensional sensor array. Hyper-spectral image data is generated by scanning the measurement sample linewise and combining the spectra of each line to a hyper-spectral image cube.

Results

One way to resolve the crosstalk due to the overlap in the emission spectra and reduce tissue autofluorescence of fluorescence measurements is spectral unmixing (SU). The method assumes that every pixel consists of a linear combination of the emission spectra of the fluorochromes. By solving the problem M(x,y,λ) = a(x,y)x1(λ) + b(x,y)x2(λ) the single components can be unmixed.

Tissue autofluorescence has a non-specific emission spectrum that causes an unwanted background intensity to be added to the intensity of every image channel and thus degrades the desired signal information. The combination of PCA, VARIMAX and ALS reduces tissue autofluorescence effectively. Compared to a standard RGB image, a reduction of tissue autofluorescence of 78% was achieved.

In a standard RGB image on average 22% of the pixels cannot be assigned unambiguously to a single class (HER2/neu,CEP17), because of the spectral overlap in the emission spectra. By applying SU methods the number of ambiguous pixels can be reduced considerably. The best result is achieved by combining PCA, VARIMAX and ALS, reducing ambiguous pixels down to 1.1%.

Conclusions

Spectral unmixing methods have been applied to hyper-spectral M-FISH images with the objective to reduce tissue autofluorescence and enhance image contrast. The combination PCA, VARIMAX and ALS reduced tissue autofluorescence by 80% for HER-2/neu spots, 64% for CEP 17 spots and 90% for cell nuclei spots. For the enhancement of image contrast the percentage of ambiguous pixels were compared. In a standard RGB image 21.8% of all pixels could not be assigned unambiguously to either CEP 17 genes, HER-2/neu genes or tissue autofluorescence. This value was reduced to 1.1%. The results show that SU is a powerful pre-processing step to improve the quality of hyperspectral M-FISH images. Subsequent (semi-) automatic analysis steps will directly benefit from a pre-processing using optimised spectral unmixing methods, thus improving both diagnostic sensitivity and reliability.

References:
[1] N. Harbeck, W. Eiermann, J. Engel, I. Funke, A. Lebeau, W. Parmanetter and M. Untch, Prognosefaktoren beim primären Mammakarzinom, In: Manual ammakarzinome Tumorzentrum München. Auser H. (Hrsg). Zuckerschwedt Verlag, 8. Auflage, pages 39-43, 2001.