Skip to main content
. 2021 Jun 24;16(6):e0253666. doi: 10.1371/journal.pone.0253666

Fig 3. Results of CNN training.

Fig 3

A) Comparison of human performance versus CNN. Pre (48.80±3.03%) and post-training (46.40±3.57%) results of five human subjects are shown compared to five separate runs of CNN training for a validation (98.18±0.33%) and a test set (96.58±0.24%). B) Representative Learning Curve of five independent CNN trainings using CPT and DMSO labels for 50 epochs. Accuracy curve for the same representative run is shown. C) Confusion matrix of CPT versus DMSO for training with highest test accuracy results. The highly accurate model led to very low false positives (65) and false negatives (52) during prediction on test set. D) Representative Learning Curve and accuracy of three independent CNN trainings using all cell lines and treatments as labels for 50 epochs. E) Confusion matrix of training with highest test accuracy results for all-versus-all analysis of test set. F) Validation accuracy results for training sets missing one cell line. The missing cell line was used as test set; testing accuracy for every run is shown in Table 2.