Figure 5.

CNN Independent Test

(A) Three independent differentiation assays were ran and images were taken as described previously. A total of 1,116 images were analyzed with both ResNet50-SA and DenseNet-SA. Confusion matrices show that both neural networks predicted with high accuracy the differentiating group.

(B) Probability plots. For each image, the neural network generates a probability for both 2i + LIF and differentiating groups. Both probabilities sum up to 1. The prediction will be based on the highest probability. Hence, all predictions are above the red horizontal lines. On the left panel, true differentiating images are represented. Except for one image in ResNet50-SA, all predictions were correct. Moreover, probabilities were very high in almost all cases. Predictions were also very high on the right panel (2i + LIF), although more variability is observed, particularly with DenseNet.

(C) Evaluation of classification accuracy during the first hour of differentiation. Differentiation was performed as previously indicated, and images were taken every 10 min during the first hour of differentiation and compared with the same number of images in the 2i + LIF condition. Images were classified according to the previously obtained ResNet50-SA training for 1 h of differentiation. Left side, confusion matrices for each time point. Right side, classification accuracy. Of note, by 20–30 min most of the differentiation images are correctly classified.