Fig. 1: CNN training to classify control and senescent cells.
From: Anti-senescent drug screening by deep learning-based morphology senescence scoring
a Representative images of input images. Input images of control and H2O2-induced senescent cells were cropped from phase-contrast microscopy images at single-cell resolution by the OpenCV-based script. Scale bar, 7.1 µm. Data are representative of over three independent experiments. b Learning curve through the CNN training. Accuracy and loss in the training data and validation accuracy and validation loss in the validation data show the process of training. c Indexes: F1 score, accuracy, precision, and recall in the final setting of training. d AUC of the ROC curve in the final setting of training. e Protocol for the evaluation of CNN generalisability. Each CNN was trained by either the images of H2O2-induced senescent cells, the images of CPT-induced senescent cells, or the mixed images of H2O2- and CPT-induced senescent cells. Newly acquired data were used as test data. In the test data, cellular senescence was induced by H2O2, CPT, or replication. Test data were evaluated by three pre-trained CNNs. f A heatmap shows the accuracy of CNN prediction in each test dataset. Three independent experiments and evaluations were conducted for each senescence induction method. g Macro-averaged accuracy for each evaluation (n = 3 independent experiments). h AUC of the receiver operating characteristic (ROC) curve in the test data evaluated by CNNs, which were pre-trained by the data from H2O2-, CPT-, or replication-induced senescent cells. Data are representative of three independent experiments. i Grad-CAM shows an important region for the prediction of healthy or senescent cells. Data are representative of three independent experiments. CNN convolutional neural network, CPT camptothecin, Rep replication.
