Abstract
BACKGROUND: For the analysis of spatio-temporal dynamics, various automated processing methods have been developed for nuclei segmentation. These methods tend to be complex for segmentation of images with crowded nuclei, preventing the simple reapplication of the methods to other problems. Thus, it is useful to evaluate the ability of simple methods to segment images with various degrees of crowded nuclei. RESULTS: Here, we selected six simple methods from various watershed based and local maxima detection based methods that are frequently used for nuclei segmentation, and evaluated their segmentation accuracy for each developmental stage of the Caenorhabditis elegans. We included a 4D noise filter, in addition to 2D and 3D noise filters, as a pre-processing step to evaluate the potential of simple methods as widely as possible. By applying the methods to image data between the 50- to 500-cell developmental stages at 50-cell intervals, the error rate for nuclei detection could be reduced to ≤ 2.1% at every stage until the 350-cell stage. The fractions of total errors throughout the stages could be reduced to ≤ 2.4%. The error rates improved at most of the stages and the total errors improved when a 4D noise filter was used. The methods with the least errors were two watershed-based methods with 4D noise filters. For all the other methods, the error rate and the fraction of errors could be reduced to ≤ 4.2% and ≤ 4.1%, respectively. The minimum error rate for each stage between the 400- to 500-cell stages ranged from 6.0% to 8.4%. However, similarities between the computational and manual segmentations measured by volume overlap and Hausdorff distance were not good. The methods were also applied to Drosophila and zebrafish embryos and found to be effective. CONCLUSIONS: The simple segmentation methods were found to be useful for detecting nuclei until the 350-cell stage, but not very useful after the 400-cell stage. The incorporation of a 4D noise filter to the simple methods could improve their performances. Error types and the temporal biases of errors were dependent on the methods used. Combining multiple simple methods could also give good segmentations.