One Shot Learning to Select Data Augmentations for Skin Lesion Classification

Abstract

Skin cancer is a highly prevalent and malignant skin disease that is usually diagnosed by visual inspection by expert doctors or dermatologists and confirmed through several supporting methods including pathological examination, medical image processing, and artificial intelligence-based techniques. Researchers and developers have been trying to leverage deep learning for the detection and classification of various forms of skin diseases. Most deep learning pipelines include data augmentation techniques to improve the overall performance of the underlying model. Selecting the right and most relevant augmentation techniques during training deep learning models is essential to achieve better performance from the underlying model. Picking up the not-so-useful augmentation methods may negatively affect the performance and computational complexity of the underlying model. In this paper, we propose a novel one-shot learning-based method to optimally pick the most relevant and useful data augmentation methods that contribute to increased model performance. The paper also focuses on designing a lightweight model, which can be easily deployed in edge networks. We train a Siamese neural network to calculate the similarity scores of 15 data augmentation techniques on images from the HAM10000 dataset. The similarity scores are used to select the most relevant data augmentation methods. Further, the experimental results confirm the increased skin lesion classification performance for the designed lightweight classification model which is trained using the augmented data, where the augmented data is generated using these selected data augmentation techniques only. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024.