Skip to main content
. 2020 Jul 3;14:171. doi: 10.3389/fncel.2020.00171

Figure 5.

Figure 5

Human-based classifier vs. CNN-based classifier. (A) The receiver operating characteristic curve for the convolutional neural networks (CNN); the area under the curve score for this classifier is 0.91. Each dot represents a single human expert predicting organoids to be from “retina” or “non-retina” class based on bright-field images. (B) Metrics comparison for human-based classifier and CNN-based. CNN showed better results on all the metrics that we measured: 0.63 vs. 0.27 ± 0.06 Matthews correlation coefficient for CNN and human, respectively; 0.84 vs. 0.67 ± 0.06 accuracy for CNN and human, respectively; 0.89 vs. 0.75 ± 0.09 F1 score for CNN and human, respectively; 0.92 vs. 0.83 ± 0.07 precision for CNN and human, respectively; 0.85 vs. 0.72 ± 0.17 recall for CNN and human, respectively.