Unsupervised meta-learning for few-shot medical image classification based on metric learning
Résumé
Due to the high costs associated with establishing large-scale medical image datasets, few-shot learning (FSL) has demonstrated remarkable potential in the field of medical image analysis. However, current FSL methods still have high labeling requirements for datasets, and FSL techniques suitable for medical scenarios need further development. Meta-learning has provided an alternative framework to address the challenging FSL setting. In this paper, we propose an Unsupervised Meta-Learning (UML) framework based on metric learning for medical image classification. We propose Mini-Batch Sampling (MBS), which optimizes the sampling strategy of UML by conducting multiple non-repeating samples of individual tasks during each training epoch. By combining a semi-normalized similarity calculation method with residual network, we obtain a stable deep learning paradigm. We extracted and reconstructed three small-sample image datasets, namely BLOOD, PATHOLOGY, and CHEST, from three publicly available medical image datasets of different types. Our proposed method outperforms current small-sample image classification methods on these three datasets, enabling the training of high-performance image classification models using small amounts of unlabeled data.