Clinically Aware Synthetic Image Generation for Concept Coverage in Chest X-ray Models

A new framework called CARPA generates synthetic chest X-ray images that preserve anatomical structure while expanding clinically meaningful concept coverage, according to research posted to the arXiv preprint server. The method aims to address a persistent limitation in medical AI: training datasets that fail to capture the full range of disease presentations seen in practice. Deep learning models for chest X-ray diagnosis are constrained by limited coverage of clinically meaningful concept combinations in publicly available training datasets [1]. While synthetic image generation has been explored to increase data diversity, existing methods rarely enforce clinical or anatomical constraints, limiting utility for improving model reliability [1]. CARPA — short for a clinically aware and anatomically grounded framework — applies targeted perturbations to clinical concept vectors while preserving anatomical structure, producing synthetic images with controlled concept insertions and deletions [1]. The research team evaluated CARPA across seven backbone architectures by fine-tuning models on synthetic subsets and testing on a held-out MIMIC-CXR benchmark [1]. Compared to prior concept perturbation approaches, fine-tuning on CARPA-generated images consistently improved precision-recall performance, reduced predictive uncertainty, and improved model calibration [1]. Structural and semantic analyses demonstrated high anatomical fidelity, strong concept alignment, and low semantic uncertainty [1]. Two expert radiologists evaluated the generated images and confirmed realism and clinical agreement [1]. The work was submitted to arXiv on 16 March 2026 and last revised on 15 June 2026 [1]. arXiv, which began on 14 August 1991, is an open-access repository of electronic preprints that are moderated but not peer reviewed, and as of November 2024 it was receiving about 24,000 articles per month [5]. The repository passed the two-million-article milestone by the end of 2021 and now spans fields including computer science, quantitative biology, and statistics [5]. Synthetic data generation has become an active area of machine learning research, paralleling broader advances in generative AI. The transformer architecture introduced in the 2017 paper “Attention Is All You Need” has become the main architecture for a wide variety of AI systems, including large language models and multimodal generative models [7]. That paper has been cited more than 250,000 times as of 2026 [7]. The CARPA framework’s emphasis on anatomical grounding distinguishes it from earlier concept perturbation methods. By ensuring that synthetic images remain anatomically faithful while inserting or deleting clinical concepts, the approach expands clinically relevant concept coverage without introducing artifacts that could mislead diagnostic models [1]. The authors argue that anatomically grounded concept perturbations enable more effective use of synthetic data, improving both performance and reliability of chest X-ray classification models and supporting safer clinical deployment [1].