The Effect of Applying Transfer Learning Approach on Medical and Non-Medical Imaging: Skin Cancer and Flower Types

Abstract

Transfer Learning is an important technique used to transfer knowledge from one pre-trained model (called source domain) to another (called destination domain), this technique improves good evaluation especially when the dataset of destination is small, transfer learning may achieve a good valuation and may not, this is depending on the common features between source and destination domain. In this work, an investigation is proposed to show the effects of using Transfer learning on medical and non-medical images. In this work, three datasets are used (International Skin Imaging Collaboration (ISIC), Human Against Machine with 10000 training images (HAM10000), and Flowers), two for skin cancer lesions as medical images and the third is flowers types, In addition, four pre-trained models are used (InceptionVersion3, Residual neural Network with 50 layers (ResNet50), Mobile network (MobileNetV2) and Extreme version of Inception (Xception). The results show that transfer learning does better using nonmedical images than medical images, and the best pre-model metrics are got from Xception model, with an accuracy of approximately 89% in non-medical images and 68% in medical images, this is because the pre-trained model is fruitful when the features are common between the source and destination domain, these common features are more available in nonmedical than medical (especially in skin lesions).