Adaptive cell nuclei detection from fluorescence images by optimizing object sizes

Although a few attempts were observed, there are high biological demands for the extraction of three dimensional cell nuclei from cellular images. Most common methods are manual and their results include subjective errors. Some other methods apply segmentation followed by centroid extraction. However, high dimensional segmentation is still unreliable, especially for biomedical images having juxtaposed cells. In this research, we propose direct method that exploits intensity distribution and object geometry to extract nuclei centroids from fluorescence microscopy images. A spatiotemporal adaptive multiscale filtering is applied for the object enhancement with noise suppression. A two-stage procedure is then followed. In Stage-1, we extract candidate centroids for the probable cell nuclei based on a local characteristic ratio (R) at every voxel. In Stage-2, an iterative procedure is proposed, which estimates approximate object radius from each candidate centroid by maximizing mean intensity gradients over successive pairs of regions in the enhanced image and removing smaller spurious objects. Investigations and quantitative analysis with a set 100 3D mouse embryo images reveal a promising performance of the proposed method in terms of average sensitivity and positional error, compared to an existing method.