Cell nuclei detection in multispectral histology images using K-means and expectation-maximisation segmentations

摘要

Histology images contain a lot of relevant information which are useful in the diagnostic (cells, cell compartments such as nuclei...). In this topic, the main goal of computer-based image analysis is to identify structures or nuclei in histology images with high accuracy and robustness. Current methods and systems based on colour images give results with a lot of errors. We suggest using multispectral imaging system with a programmable light source (PLS). With the new acquisition system, a 3-band colour image (MS3), a 5-band multispectral image (MS5), a 10-band multispectral image (MS10) and a 25-band multispectral image (MS25) from 450 nm to 700 nm are acquired. After the acquisition, two unsupervised segmentation methods are applied: the expectation-maximisation (EM) and the K-means (KM). Firstly, each band is segmented separately; secondly a fusion of bands is used. A comparison has been drawn between the two segmentation methods. The results show a small superiority of EM segmentation against KM segmentation. It is also noted that the fuse of selected bands from MS5 ensures the best F-measure of cell nuclei detection.