LSNet: a deep learning based method for skin lesion classification using limited samples and transfer learning.

When analyzing skin lesion image data using deep learning, the lack of a sufficient amount of effective training data poses a challenge. Although transfer learning can alleviate the problem of a small amount of data, the difference between the source data and the target data makes the tranfer learning missing some key knowledge. It is important to find the key knowledge neglected by transfer learning. This paper argues that this key knowledge is contained in challenging samples. We propose a novel method named as Limited Samples Network (LSNet) to search challenging samples and strengthen the learning of them. Specifically, LSNet utilizes patch-based structured input and employs pseudoinverse learning autoencoder to quickly obtain position-sensitive loss. Challenging samples can be obtained by searching for position-sensitive loss. Subsequently, challenging samples-augmented transfer learning is employed to enhance the classification performance of deep learning models on skin lesion datasets with limited samples. We carry comparison experiment with the existing state-of-the-art method. Experiments are carried out on the ISIC 2017, ISIC 2018 and ISIC 2019 skin lesion datasets. The results demonstrate that our training strategy significantly improves the vanilla transfer learning procedure for different types of pre-trained DCNNs. In particular, our method achieves state-of-the-art performance on different skin lesion datasets without using any extra training data. [ABSTRACT FROM AUTHOR]