Efficient Transfer Laerning Used in the Classification of Gastroscopic Images with Small Dataset

Deep learning (DL) has gained success in image classification. However, DL always needs a large number of labeled training data which are not usually available for gastroscopic image analysis. Therefore, the pre-trained convolutional neural networks (CNNs) models are important. The transfer learning (TL) method could leverage the pre-trained model in natural images to gastroscopic images. In this paper, we have studied different TL methods on small dataset of gastroscopic images. We have fine-tuned the current deep CNNs: EfficientNet, ResNet and DenseNet on three kinds of gastroscopic images from kvasirv2 dataset. The results demonstrated that the pre-trained models of current deep CNNs can get considerable classification results on gastroscopic images with small dataset.