Figure 3 | Scientific Reports

Figure 3

From: Construction of a system using a deep learning algorithm to count cell numbers in nanoliter wells for viable single-cell experiments

Figure 3

Defining ambiguous samples increases prediction accuracy. Ambiguous samples were defined by their highest outputs being lower than a threshold, and subsequently assessed by trained technicians. (a) A box plot comparison of log transformed highest outputs of concordant (7835) and discordant (133) samples. The results clearly indicate that discordant samples had much lower highest outputs as compared with concordant samples (P < 2 × 10−16, Mann-Whitney’s U test). (b–d) Change in the concordance rate between the results from the trained CNN and the trained technicians by changing the threshold value for the highest output from CNN for each sample. The samples with the highest outputs lower than the threshold value were classified as ambiguous samples. In this analysis, the ambiguous samples are assumed to be subsequently judged by technicians and all will become concordant. When the proportions of the ambiguous samples are 0 (b), 1 (c), and 5% (d), the concordance rates will become 98.3%, 99.0%, and 99.5% respectively.

Back to article page