Abstract

Multi-label Chest X-Ray classification is crucial for intelligent diagnosis, yet existing algorithms usually ignore lesion-scale heterogeneity and attribute conditioned label dependencies, limiting their clinical generalizability. To address these issues, this paper proposes MSASG, a multi-label Chest XRay image classification algorithm based on Multi-Scale and Attribute-aware Semantic Graph which enhances discriminative power and semantic consistency. Firstly, a Multi-scale Feature Partitioning and Reconstruction method is proposed to capture lesion patterns at different scales. Secondly, a Label guided Multi-scale Semantic Alignment method is proposed to improve visual–semantic alignment by integrating label embeddings into feature extraction and using a Transformer to model high-order cross-modal dependencies.
Finally, an Attribute-aware Graph Convolutional Network method is proposed to construct attribute-specific label co-occurrence matrices and dynamically select relevant structures during inference, enabling personalized characterization of label dependencies. Experiments on ChestX-ray14 and CheXpert show that MSASG outperforms state-of-the-art methods in recognizing complex lesion co-occurrence and adapting to heterogeneous populations