E2CL: An Efficient and Effective Classification Learning for Pneumonia Detection in Chest X-rays

The goal of pneumonia detection in chest X-ray images is to accurately and efficiently identify cases of pneumonia.xisting models tend to suffer from insufficient generalization and overhead when dealing with large-scale data. To alleviate above problems, this paper proposes an Efficient and Effective Classification Learning (E2CL) framework built upon the self-attention mechanism of Swin Transformer. Specifically, a data augmentation strategy is initially employed to enrich the diversity of the dataset and improve the generalization of the model. Based on the expanded data, the MSG Token and the hard distillation network with teacher-student guidance are then introduced to reduce the complexity of the Swin Transformer and ensure its performance. The MSG Token facilitates effective information transfer between local windows, thereby avoiding the computational cost associated with traditional SW-MSA. The student model is trained using hard labels of the teacher model, enabling it to learn more high-quailty feature representation. Additionally, this manner of training with hard labels helps mitigate overfitting caused by limited data availability. Extensive experimental results show that the E2LC model in the Chext X-Ray dataset achieves an accuracy of 91.7% on the three-classification task and 93.06% on the two-classification task, which significantly outperforms multiple existing state-of-the-art models. The effectiveness of each component is further validated through ablation experiments with data augmentation, MSG Token and distillation methods.