Multispectral imaging is becoming a key technique for biomedical research, but the crosstalk between autofluorescence and fluorescent material severely affects the interpretation of fluorescence images. Spectral unmixing is an effective technique for removing autofluorescence and separating fluorescent targets in multispectral fluorescence imaging. However, the effectiveness of most methods of spectral unmixing has a strong relationship with the noise in the image. In this work, we propose a multispectral fluorescence unmixing method based on a priori information to obtain the pure spectra and their corresponding abundance coefficients in the images. First, the obtained multispectral image is segmented into several superpixels using a superpixel segmentation method, and then the relative pure spectra are extracted using a spectral extraction algorithm on the superpixels. Since the autofluorescence distribution is spread over the whole body, the extracted spectra in which the autofluorescence can be considered as pure spectra are used as a priori knowledge for unmixing. The pixel spectral data that are similar to the set of relatively pure spectra are selected as the pure spectral candidate set. Then the pure spectra can be obtained using the Non-negative matrix factorization method with prior knowledge(NMF-upk). Finally, the abundance corresponding to each spectral feature can be obtained through the least square method. The proposed unmixing method is tested on simulated data and the results show that our unmixing algorithm outperforms other methods.
|