Robust Cross-Center Chest X-ray Anomaly Detection: via Dynamic Memory and Transformer Inpainting (Preprint)

BACKGROUND Robust chest X-ray (CXR) anomaly detection that generalizes across multiple clinical centers is notoriously difficult because variations in X-ray generators, image post-processing pipelines and patient demographics induce substantial domain shifts. In routine deployment, AI systems are expected to operate “out-of-the-box” at new hospitals that cannot share proprietary images for fine-tuning. OBJECTIVE To develop and validate a Dynamic-Memory Autoencoder (DMA) that maintains high performance on unseen multi-center CXR data without any domain-specific fine-tuning. METHODS We propose a three-component architecture: (i) a dynamic memory network whose slots are written to and pruned on-the-fly according to gated residuals; (ii) a transformer-based adaptive in-painting head that enforces long-range anatomical consistency; (iii) a lightweight teacher–student distillation scheme that stabilizes training. DMA was first pre-trained on a heterogeneous synthetic CXR corpus that emulates a broad spectrum of normal anatomy and common pathologies. External validation used 300 real-world CXRs collected between 2023 and 2025 from five scanners. RESULTS On the pre-training corpus, DMA achieved 84.7% AUC and 80.05% accuracy, surpassing four state-of-the-art baselines by up to 14 AUC points. In the multi-center external validation, DMA preserved 83.4% AUC and 80.23% accuracy without any domain-specific fine-tuning—degradation of only 1.3 pp and 0.1 pp, respectively. Qualitative analyses showed that the dynamic memory rapidly expanded to unfamiliar textures, while the transformer head preserved sharp reconstructions of healthy tissue and produced high-contrast residuals over lesions. CONCLUSIONS DMA offers a practical, scalable solution for multi-center CXR triage and could help alleviate radiologists’ workload in resource-limited settings. CLINICALTRIAL Not applicable.