Transfer learning with multiple pre-trained network for fundus classification

Abstract

Transfer learning (TL) is a technique of reuse and modify a pre-trained network. It reuses feature extraction layer at a pre-trained network. A target domain in TL obtains the features knowledge from the source domain. TL modified classification layer at a pre-trained network. The target domain can do new tasks according to a purpose. In this article, the target domain is fundus image classification includes normal and neovascularization. Data consist of 100 patches. The comparison of training and validation data was 70:30. The selection of training and validation data is done randomly. Steps of TL i.e load pre-trained networks, replace final layers, train the network, and assess network accuracy. First, the pre-trained network is a layer configuration of the convolutional neural network architecture. Pre-trained network used are AlexNet, VGG16, VGG19, ResNet50, ResNet101, GoogLeNet, Inception-V3, InceptionResNetV2, and squeezenet. Second, replace the final layer is to replace the last three layers. They are fully connected layer, softmax, and output layer. The layer is replaced with a fully connected layer that classifies according to number of classes. Furthermore, it's followed by a softmax and output layer that matches with the target domain. Third, we trained the network. Networks were trained to produce optimal accuracy. In this section, we use gradient descent algorithm optimization. Fourth, assess network accuracy. The experiment results show a testing accuracy between 80% and 100%.

Prediction of Molecular Properties Using Molecular Topographic Map

“…In the future work I intend to develop a method to extract chemical information from the important regions on MTM for predicting molecular property. In this study, although simple and basic CNN architecture was used, the prediction performance of CNN using MTM is expected to be improved by using transfer learning via multiple pretrained networks, such as AlexNet, VGG19, RsNet101, GooLeNet, and Inception-V3 [ 46 ]. I believe that MTM embedded molecular structure information can serve as a valuable method of molecular representation for drug discovery.…”

Convolution neural network model for fundus photograph quality assessment

“…When huge data samples (i.e., photographs) are involved, there will be some parameters in CNN layers to be considered [40], [41]. However, small datasets will cause overfitting [40].…”

A Preliminary Intergenerational Photo Conversation Support System based on Fine-tuning VGG16 Model

“…It used rotation, skewing and elastic distortion augmentation methods for images and then used a pre-trained CNN model as feature extractor and SVM as a category classifier. This technique has been applied in many fields and has shown better accuracy than traditional convolutional networks [25][26][27]. In this paper after we augmented the training set, we combined transfer learning and fine-tuning and produced five models (Fig.…”

“…The reason for this result is that models provided in this paper reduced the irrelevant information of photos. From the perspective of models, the main focus of preventing overfitting should be the entropy capacity of the model, or how much information the model can store [25,27]. There are different ways to modulate entropic capacity, the main one is the choice of the number of parameters in your model, i.e.…”