DIAN: Device Invariant Attention Network for Chest X-ray Classification

Deep learning has significantly advanced automated chest X-ray interpretation, yet models remain vulnerable to spurious correlations with visual confounders such as support devices. We propose the Device Invariant Attention Network (DIAN), which couples spatial self-attention with domain-adversarial learning to disentangle disease-relevant signals from device-induced artifacts. DIAN employs an attention module that highlights pathology-relevant regions for prediction while simultaneously identifying device-related areas, which are suppressed via an adversarial domain classifier trained with device presence as the domain label. We further argue that standard ranking metrics are insufficient in the device-invariant setting and therefore complement AUC with threshold-based metrics-F1 score, precision, and recall-to assess whether predictions remain stable when device cues are present or absent. This evaluation exposes cases where the decision boundary is shifted by support devices and provides a more reliable assessment of device-independent disease prediction. Evaluated on the CheXpert benchmark, DIAN improves generalization on device-sensitive pathologies and attains a mean area under the ROC curve (AUC) of 0.909, achieving competitive result compared to prior baselines.