Scalable empirical methods for evaluating and optimizing paraphrase robustness in multi-center chest X-ray dataset generalization.

Abstract

The deployment of vision-language models (VLMs) in medical imaging, particularly for chest X-ray analysis, necessitates robust generalization across diverse clinical settings. However, paraphrastic variations in clinical language can significantly degrade model performance, posing a critical challenge for reliable multi-center deployment. This manuscript introduces scalable empirical methods for evaluating and optimizing paraphrase robustness in VLMs applied to multi-center chest X-ray datasets. We propose a novel framework, Paraphrase Robustness Assessment and Calibration (PRAC), which integrates a lightweight paraphrase generation engine with a distributional robustness objective. Our methodology formalizes paraphrase robustness as a min-max optimization problem, leveraging a multi-objective loss that combines contrastive alignment and adversarial training. We introduce a new evaluation metric, Paraphrase Consistency Score (PCS), which quantifies the semantic stability of model outputs under paraphrastic perturbation. Through extensive experiments on MIMIC-CXR, PadChest, and VinDr-CXR, we demonstrate that our approach significantly improves cross-center generalization, achieving up to 12% improvement in PCS over baseline models. We also provide a comprehensive analysis of failure modes, showing that models are particularly sensitive to clinical terminology variations and negation handling. Our work provides a scalable, robust solution for deploying VLMs in real-world medical settings, with implications for clinical decision support and multi-institutional collaborations.