cs.CVMar 16, 2026

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

Authors: Amy Rafferty, Rishi Ramaesh, Ajitha Rajan

Organizations: University of Edinburgh, UK · NHS Lothian, UK

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Deep learning models for chest X-ray diagnosis are constrained by limited coverage of clinically meaningful concept combinations in publicly available training datasets. 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. We propose CARPA, a clinically aware and anatomically grounded framework for synthetic chest X-ray generation that applies targeted perturbations to clinical concept vectors while preserving anatomical structure. By producing anatomically faithful synthetic images with controlled concept insertions and deletions, CARPA expands clinically relevant concept coverage. We evaluate CARPA across seven backbone architectures by fine-tuning models on synthetic subsets and testing on a held-out MIMIC-CXR benchmark. Compared to prior concept perturbation approaches, fine-tuning on CARPA-generated images consistently improves precision-recall performance, reduces predictive uncertainty, and improves model calibration. Structural and semantic analyses demonstrate high anatomical fidelity, strong concept alignment, and low semantic uncertainty. Evaluation by two expert radiologists further confirms realism and clinical agreement. Together, these results show 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.

Figures & tables

Figure 1: Example synthetic images generated by CoRPA and CARPA. Image generation prompts: (1) Pneumonitis, (2) Effusion
Figure 2: Comparison of label (a) and concept (b) distributions between the original dataset (N=39921), CoRPA (N=10000) and CARPA (N=10000). Values are proportions of respective datasets.
  ConvNeXt-BConvNeXt-SDenseNetEfficientResNet50ResNet101ResNeXt101
BaselineAUROC0.9470.9400.9220.9430.9130.9260.949
 AUPRC0.8640.8380.7720.8390.7190.7860.869
 F10.8350.8020.7340.7970.6770.7400.844
Finetuned CARPAΔ\Delta AUROC↓\downarrow 0.007↓\downarrow 0.008↓\downarrow 0.008↓\downarrow 0.010↑\uparrow 0.003↓\downarrow 0.007↓\downarrow 0.009
(Ours)Δ\Delta AUPRC↑\uparrow 0.048↑\uparrow 0.052↑\uparrow 0.110↑\uparrow 0.026↑\uparrow 0.087↑\uparrow 0.058↓\downarrow 0.003
 Δ\Delta F1↑\uparrow 0.087↑\uparrow 0.092↑\uparrow 0.096↑\uparrow 0.019↑\uparrow 0.102↑\uparrow 0.033↑\uparrow 0.002
Table 1: Comparison of multi-label classification performance across seven backbone architectures. We report macro-averaged AUROC, AUPRC, and F1 for baseline models trained on a subset of MIMIC-CXR, and performance deltas (increase ↑\uparrow or decrease ↓\downarrow ) for corresponding variants fine-tuned on the CARPA (Ours) or CoRPA synthetic datasets. All metrics are evaluated on the same held-out test split (N=4008), with class-specific decision thresholds calibrated on the validation set when computing F1. Performance increases shown in bold .
  ConvNeXt-BConvNeXt-SDenseNetEfficientResNet50ResNet101ResNeXt101
BaselineEntropy0.0860.1350.2250.1230.2650.2010.043
 ECE0.0450.0630.0930.0620.1140.0850.040
Finetuned CARPAΔ\Delta Entropy↓\downarrow 0.012↓\downarrow 0.022↑\uparrow 0.026↑\uparrow 0.024↓\downarrow 0.105↓\downarrow 0.075↓\downarrow 0.002
(Ours)Δ\Delta ECE↓\downarrow 0.012↓\downarrow 0.008↓\downarrow 0.014↓\downarrow 0.010↓\downarrow 0.031↑\uparrow 0.006↓\downarrow 0.009
Finetuned CoRPAΔ\Delta Entropy↑\uparrow 0.057↑\uparrow 0.055↑\uparrow 0.062↑\uparrow 0.071↑\uparrow 0.032↑\uparrow 0.035↑\uparrow 0.062
 Δ\Delta ECE↑\uparrow 0.094↑\uparrow 0.097↑\uparrow 0.080↑\uparrow 0.125↑\uparrow 0.051↑\uparrow 0.083↑\uparrow 0.154
Table 2: Uncertainty and calibration analysis across seven backbone architectures. For each model, we report predictive entropy and Expected Calibration Error (ECE) for the baseline model trained on the MIMIC-CXR subset, as well as performance deltas (increase ↑\uparrow or decrease ↓\downarrow ) for variants fine-tuned on the CARPA (Ours) and CoRPA synthetic datasets. All metrics are computed on the same held-out test split (N=4008), with values reported as macro-averages across labels. Uncertainty reductions shown in bold .
 Perturbation TypeSSIM ( Mean±sdMean\pm sd )
CARPAIntra-class0.748±0.0300.748\pm 0.030
(Ours)Insertion0.751±0.0320.751\pm 0.032
 Deletion0.749±0.0300.749\pm 0.030
 Overall0.750±0.031\mathbf{0.750\pm 0.031}
CoRPAIntra-class0.304±0.0630.304\pm 0.063
 Outer-class0.229±0.1020.229\pm 0.102
Table 3: Structural similarity (SSIM) between original chest X-rays and synthetic images generated by CARPA (ours) and CoRPA. Higher SSIM indicates greater preservation of anatomical structure.
 JaccardHammingUFP=1−PrecisionU_{\text{FP}}{=}1{-}\text{Precision}UFN=1−RecallU_{\text{FN}}{=}1{-}\text{Recall}
CARPA (Ours)0.609±0.1490.609\pm 0.1490.092±0.0430.092\pm 0.0430.489±0.2110.489\pm 0.2110.287±0.1140.287\pm 0.114
CoRPA0.168±0.3270.168\pm 0.3270.113±0.0680.113\pm 0.0680.799±0.3590.799\pm 0.3590.595±0.4880.595\pm 0.488
Table 4: Semantic concept uncertainty evaluation for CARPA (Ours) and CoRPA subsets using the LLaVA-Rad image-to-text pipeline. Values are reported as mean ±\pm standard deviation over 10,000 images.
 TaskResponseR1R2
CARPAReal / SyntheticReal94%88%
(Ours) Synthetic6%12%
 Clin. AgreementFully Agree61%59%
  Partial Agree22%21%
  Disagree17%20%
CoRPAReal / SyntheticReal86%62%
Table 5: Expert evaluation of synthetic chest X-rays generated by CARPA (N=100) and CoRPA (N=50). Two radiologists (R1, R2) independently assessed images for realism (Real / Synthetic) and agreement with proposed clinical findings (Clin. Agreement).

Explore similar work

CardsList

  1. CheXanatomy: Anatomy-Aware Vision-Language Modeling for Chest Radiographs

    Jun 7, 2026Sergios Gatidis, Curtis Langlotz, Christian BluethgenRadiology VLMsMedical Image Segmentation

    Stanford Center for Artificial Intelligence in Medicine and Imaging, Stanford University, Palo Alto, CA, USA · Department of Radiology, Stanford University, Stanford, CA, USA

  2. Scaling Generative Foundation Models for Chest Radiography with Rectified Flow Transformers

    Jun 17, 2026Fabio De Sousa Ribeiro, Emma A. M. Stanley, Charles Jones +7Medical Imaging Foundation ModelsMedical Image Synthesis

    Imperial College London · Causality in Healthcare AI Hub · Cleveland Clinic London +4

  3. 2D Versus 3D Diffusion for In Silico Training of Interventional X-ray AI Models

    Jun 19, 2026Sampath Rapuri, Jeremy Ko, Benjamin D. Killeen +2Synthetic Data AugmentationMedical Imaging

    Johns Hopkins University, Baltimore, MD 21218, USA