CONCATENATION OF EFFICIENTNETB7 AND RESNET50 MODELS IN THE TASK OF CLASSIFYING OPHTHALMOLOGICAL DISEASES OF DIABETIC ORIGIN

Diagnosing diabetic eye diseases by doctors using medical equipment requires significant resources. It is advisable to use automated tools. Using combinations of models improves classification accuracy. The features of the architectures of convolutional neural networks EfficientNetB7 and ResNet50 are presented. The creation of a neural network model by concatenating the EfficientNetB7 and ResNet50 models is justified. Transfer learning is applied. The GlobalAveragePooling2D layer is added to each model. The models are combined using the Concatenate layer. The Flatten layer is added to the resulting model to convert the vector into a one-dimensional array. Two Dropout layers are added to prevent overtraining. Two Dense layers with 512 and 256 neurons and the ReLU activation function are added for nonlinear data transformation and abstract feature extraction. A Dense layer with 4 neurons and the softmax activation function is added to determine the image class. l2-regularization is used in all Dense layers. The developed neural network model was applied to process a dataset of 4 classes: cataract images, diabetic retinopathy images, glaucoma images, and healthy retina images. The model is compiled using the Adam optimizer, the categorical cross-entropy loss function. The callback functions ModelCheckpoint, LearningRateScheduler, EarlyStopping, and ReduceLROnPlateau are used to adjust the learning rate. The validation accuracy of the model is improved by augmentation (horizontal and vertical flipping), using l2-regularization, Dropout, and adjusting the callback functions. The training lasted 30 epochs. The best validation accuracy of 97.39% was achieved at the 29th epoch. The best value of the validation function 0.4323 was achieved at the 30th epoch. The proposed neural network model outperforms the accuracy indicators of models proposed in similar studies. The model can be applied to disease detection and classification tasks.