Patch-Based Multi-Level Attention Mechanism for Few-Shot Multi-Label Medical Image Classification

Few-shot learning stands as a prominent trend in the field of computer vision, with substantial applications in vision tasks such as image classification and semantic segmentation. It has gained popularity due to its potential to reduce the demand for computer resources and its ability to lessen dependence on large datasets. However, generating high-performance models becomes challenging since this approach must generalize only from a limited set of samples. This challenge is particularly evident in multi-label medical image classification, where overlapping labels and obscure characteristics within specific image regions impede the generalization capabilities of few-shot learning. This paper proposes a patch-based strategy with a multi-level attention mechanism. Our approach employs patch-based methods with multi-level attention to segment regions with overlapping information in images, thereby facilitating the extraction of crucial feature data. Experimental results reveal that the patch-based technique can help multiple models achieve greater classification performance across various datasets, demonstrating that the strategy effectively addresses the challenges inherent in multi-label classification.