Fig. 3: Optimization of CNN model.

a, b Training and validation accuracy varying the size of images. c test accuracy varying training image size and 224 × 224 pixel case is selected because the model is aiming for the ImageNet dataset, which has 224 × 224 sized images. d examples of cropped training images with scale bar length of 20 µm, and e, f training and validation accuracies as a function of epochs for different number of training images.