Fig. 4
From: Neuron ID dataset facilitates neuronal annotation for whole-brain activity imaging of C. elegans
An automatic annotation method and evaluation. a The outline of the atlas generation method. b The outline of the automatic annotation method. The schemes of bipartite graph matching and majority voting are shown. c Error rates of the automatic annotation method for the animals in the neuron ID dataset. The names of the cells were estimated based on their positions. The error rate was calculated as 1 – (Ncorrect)/(Nannotated) for each animal, where Nannotated is the number of human-annotated cells (ground truth) and Ncorrect is the number of cells whose annotation by the algorithm was correct. Cells un-annotated by human were not included in the calculation of error rate. The rank R indicates that it is considered correct if the correct annotation appeared in the top R estimations by the algorithm. The error rates were evaluated by cross-validation, and mean ± standard deviation over well-annotated six animals is shown. d Error rates of the automatic annotation method for the strain JN3039 that expresses the fluorescent landmarks. The names of the cells were estimated based on their positions with or without the expression of landmark promoters. Mean ± standard deviation over 15 animals is shown. e The automatic annotation method was integrated in the graphical user interface roiedit3d that enables feedback between automatic and manual annotations. f The effect of manual correction on the error rate of automatic annotation. A wrong annotation of a cell in rank 1 estimation for JN3039 (see Fig. 4d) was corrected, and the automatic annotation method was performed by using the correction information. This step was repeated sequentially