Figure 3

Neural Networks are able to discriminate between drug sensitive and non-sensitive spheroids. (A) Graphical representation of the convolutional neural network (CNN). Images are convoluted with kernel matrices in order to reduce complexity but to keep the image information. Fully connected layers are then used to annotate an input image to a given classification. (B) Averaged training accuracy (‘acc_avg’, blue line) and validation accuracy (‘val_acc_avg’, orange line) of the CNN in a five-fold cross validation for 50 epochs. (C) The level of precision of the image classification by the CNN as calculated by a five-fold cross validation for the three individual categories “unaffected”, “mildly affected” and “affected. Numerical values are the means. (D) The level of recall of the image classification by the CNN as calculated by a five-fold cross validation for the three individual categories “unaffected”, “mildly affected” and “affected. Numerical values are the means. (E) The F1-score as a function from precision (C) and recall (D) for the three individual categories of spheroid viability: F1 = 2 × (precision × recall)/(precision + recall). Numerical values are the means.