Unbalanced few-shot classification based on ME-MAML framework

Aiming at the few-shot problem of lack of sufficient samples and uneven samples, which is common in medical image classification tasks, this paper proposes a meta-learning framework ME-MAML (Medical ModelAgnostic Meta-Learning) for few-shot medical image classification tasks. The framework aims to train a set of parameters with good generalization performance for the multi-class medical image classification task by learning the base class of medical images, and complete the generalization for the new class of medical image classification task by learning a small number of new classes of medical images. ME-MAML improves the MAML (Model-Agnostic Meta-Learning) meta-learning framework from the aspects of the feature extraction network, the loss function and the algorithmic framework, which makes the framework more suitable for few-shot medical image classification tasks. The experimental results demonstrate that for few-shot medical image classification task, the framework has better classification performance compared to supervised learning frameworks as well as advanced meta-learning frameworks.