AGFR-Net: anatomy-guided feature refinement for robust multi-label thoracic disease classification from chest X-rays

Detecting thoracic diseases from chest X-ray images remains difficult because many abnormalities appear with subtle visual patterns and are often obscured by overlapping anatomical structures. Another major challenge is the highly uneven distribution of disease labels in large chest X-ray datasets. Recent deep learning methods have explored the use of attention mechanisms and anatomical priors, but many existing approaches still combine these components through parallel fusion or single-stage recalibration instead of progressively refining feature representations across multiple learning stages. In this work, we propose AGFR-Net, a hierarchical anatomy-guided feature refinement framework that combines anatomical localization with disease-specific representation learning through a sequential coarse-to-fine refinement strategy. The model first employs soft lung-region priors, derived from a quantitatively validated pre-trained segmentation network, to guide feature extraction toward anatomically relevant regions. These anatomy-guided representations are subsequently refined through a cascaded attention mechanism consisting of channel-wise, spatial, and element-wise refinement stages to progressively refine disease-relevant features. To address severe class imbalance, we adopt a hybrid loss formulation that combines focal modulation with class-weighted binary cross-entropy (WBCE). The segmentation model used for anatomical guidance was quantitatively evaluated on public lung segmentation benchmarks and achieved strong Dice coefficient and Intersection over Union performance, supporting the use of soft anatomical priors within the proposed framework. We evaluate AGFR-Net on the NIH ChestX-ray14 dataset under controlled experimental settings and further assess cross-dataset performance trends using CheXpert without additional fine-tuning. The proposed framework was associated with improved performance across multiple thoracic disease categories, particularly for relatively underrepresented conditions, supported by controlled ablation studies and bootstrap-based statistical analysis. We also noticed that the performance of the models differed in severe pathological cases, depending on the quality of segmentation, dataset features and resolution of the input image. These observations suggest a possible beneficial effect of gradually incorporating anatomical information when refining features, in order to make the classification of chest X-ray more robust in heterogeneous clinical imaging settings.