Title: PaCX-MAE: Physiology-Augmented Chest X-Ray Masked Autoencoder

Abstract: Although clinical diagnosis frequently necessitates the integration of imaging data with physiological metrics, most deployed models are limited to unimodal inputs. To address this, we introduce PaCX-MAE, a cross-modal distillation framework that embeds physiological priors into chest X-ray (CXR) encoders during training while maintaining strict unimodal operation at inference time. This approach enhances in-domain masked autoencoding by incorporating a dual contrastive-predictive loss, thereby aligning CXR representations with corresponding embeddings from ECG and laboratory data. Our comprehensive evaluation across nine distinct benchmarks reveals consistent performance gains over domain-specific MAE models, with notable improvements in physiology-dependent tasks, including a +2.7 AUROC increase on MedMod and a +6.5 F1 score improvement on VinDr. The methodology demonstrates high label efficiency within the 1% data regime and maintains anatomical accuracy, matching the performance of standard MAE on segmentation tasks. Furthermore, zero-shot evaluations and attention analyses indicate that PaCX-MAE effectively learns to focus on physiological cues, such as the cardiac silhouette, which are typically overlooked in conventional visual pretraining.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC