Abstract:

Nucleus automatic detection is not only an important part of pathological image analysis, but also one of the primary bottlenecks to improve the analysis accuracy of the technology. The reason lies in that the pathological slice exhibits uneven layering and staining, nucleus crowding or overlapping. In order to improve the accuracy of the nucleus detection, we design a nucleus model of multi-curvature contour. The nucleus contour features are extracted by multicurvature orientation energy filter. Then the contour features, along with the cell nucleus ground truth, are put into a boosting algorithm, which creates a pixel classifier to classify pixels into foreground and background. In the end, the mean-shift algorithm produces the confidence coefficiency of the nucleus for each location. Experimental results show that in comparison with other state-of-art cell nuclei detection methods, our method is more robust to conditions such as biological variations, different staining and illumination conditions, and touching or partial occlusions.

Key words: nucleus detection;multi-curvature;pattern recognition;machine learning