Deep learning

We used two Convolutional Neural Network (CNN) models: the Neural Network Console (NNC; SONY) and the Keras-Tensorflow backend (Google). The VGG16 model is comprised of five blocks with three fully connected layers. Each block includes the convolutional layers followed by a max-pooling layer. A flattening of the output from block 5 resulted in two fully connected layers. In the current study with the Keras model, we adopted and implemented VGG16, a pre-trained CNN architecture that won the ImageNet Large Scale Visual Recognition Challenge (ILSVRC). The Keras model is fine-tuned by transfer learning. The NNC model resembles the VGG16 and was trained from random initialization. We compared the performance of Keras which uses conventional programming to that of NNC which is simpler in function but does not require programming knowledge.

The original en face image (550 x 550 pixels) was cropped to remove the optic disc and resized to 224 X 224 pixels. The cropped and resized image was used for both models. The entire 100 en face images were randomly divided into 80 for training and 20 for validation. We calculated the correct answer rate during the validation process. For this, we divided it into five groups of 20 eyes each, trained the remaining 80 eyes in each group, and then calculated the correct answer rate for each group by validation with 20 eyes. Thus, each group was trained and validated in the same process for a total of five times as a cross-validation of the method.

The NNC models were trained with a batch size of 20, epochs of 100, and with Adam optimization (learning rate 0.001). The Keras models were trained with a batch size of 10, epochs of 100, and with SGD optimization (learning rate 0.001). The batch size was adjusted for each model to avoid over-training. SGD optimization was used in the Keras model to implement the Heatmap.