A Study of Imbalanced Medical Image Classification Based on Self-Supervised Learning Incorporating Attention Mechanisms

The distribution of medical image datasets is often unbalanced, which is usually because the population with a specific disease is relatively small compared to the normal population, and it is very important to solve the problem of class imbalance. However, if the class imbalance samples are directly given to the model for learning, a few classes of specific diseases are seriously underrepresented in the datasets, resulting in overfitting and poor performance in the model learning process. In this paper, SSP-ECABM algorithm is proposed. To solve the problem that self-supervised learning algorithm has insufficient representation ability in medical images, ECA attention mechanism is introduced into self-supervised network structure to effectively improve the feature learning ability of self-supervised network model. On the basis of the improved self-supervised network, in order to overcome the inherent label bias problem, unlabeled self-supervised pre-training (SSP-ECA) is carried out, and then labeled supervised training is carried out downstream of the self-supervised network. To solve the problem of class imbalance, this paper proposes to use unbalanced classifier to replace the traditional linear classification strategy in the self-supervised downstream, and use balanced-Mixup (BM) unbalanced learning module to generate more Balanced training data distribution in the top-level module of the self-supervised network, so as to avoid the underfitting of a few classes.