Improving the detection of cell centroids from fluorescence images by adaptive filtering

Automated nuclei detection and counting from higher dimensional fluorescence images is crucial for many biomedical applications. Common methods either use segmentation followed by centroid extraction or directly extract nuclei centroids. Although direct methods are desirable in order to avoid complex segmentation, most of these methods disregard the temporal variations of nuclei sizes due to cell division while performing enhancement filtering. As a result, accuracy of detection becomes low. In this study, we propose a modified direct method that includes time-adaptive filter masks for improving detection accuracy. We employ cube function as base filter for the multiscale enhancement. A data-driven technique which capitalizes the approximate object volume is proposed to compute normalized-volume-ratio function for the time-series data. This function correlates the decreasing object sizes on average and generates filtering parameters at every time point based on some base parameters, empirically selected. A three-stage procedure is then followed. We extract candidate centroids from the enhanced image using a characteristic ratio at every voxel. Stage-1 centroids are then refined by analyzing the shapes of the intensity profiles of enhanced image (Stage-2). An iterative procedure is finally adopted to combine fragmented nuclei (Stage-3). Investigations and quantitative analysis with a set of 100 3D mouse embryo images reveal a promising achievement of the technique presented in terms of average F-measure (92.64% by proposed), when compared with two existing methods (87.62% and 89.69 %).