SynthCheX-75K

SynthCheX-75K is a synthetic dataset developed from Sana (0.6B) [1], a highly-capable text-to-image model, fine-tuned on chest radiographs. The dataset contains 75,649 high-quality samples after a strict filteration process. This dataset is released with the paper: CheXGenBench: A Unified Benchmark For Fidelity, Privacy and Utility of Synthetic Chest Radiographs.

Filtration Process for SynthCheX-75K

Generative models can lead to both high and low-fidelity generations on different subsets of the dataset. In order to keep the sample quality high in SynthCheX-75K, a stringent filtration process was adopted using HealthGPT [2], a highly-capable medical VLM with advanced understanding, reasoning and generation.

The VLM was provided with the following meta-prompt to classify the quality of each generated sample.

META PROMPT: "You are an expert radiologist and medical annotator. Your task is to assess the quality of an image given its description and classify the image as either 'High Quality', 'Medium Quality', or 'Low Quality'. Keep your responses limited to only these three options. If the image is not relevant to the description, respond with 'Not Relevant'."

After quality label assignment, images with "Low Quality" and "Not Relevant" labels were removed from the dataset leading to 75,649 samples of high-quality radiographs with pathological annotations.

Dataset Usage

from datasets import load_dataset
dataset = load_dataset("raman07/SynthCheX-75K", trust_remote_code=True)
print(dataset['train'][0])

{'image': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=512x512>,
 'filename': 'SyntheticImg_02aa804e-bde0afdd-112c0b34-7bc16630-4e384014.png',
 'prompt': 'There is no focal consolidation, pleural effusion or pneumothorax. Bilateral nodular opacities that most likely represent nipple shadows. The cardiomediastinal silhouette is normal. Clips project over the left lung, potentially within the breast. The imaged upper abdomen is unremarkable. Chronic deformity of the posterior left sixth and seventh ribs are noted.',
 'labels_dict': {
    'Atelectasis': 0,
    'Cardiomegaly': 0,
    'Consolidation': 0,
    'Edema': 0,
    'Enlarged Cardiomediastinum': 0,
    'Fracture': 0,
    'Lung Lesion': 0,
    'Lung Opacity': 0,
    'No Finding': 1,
    'Pleural Effusion': 0,
    'Pleural Other': 0,
    'Pneumonia': 0,
    'Pneumothorax': 0,
    'Support Devices': 0
    }
}

CheXGenBench: A Unified Benchmark For Fidelity, Privacy and Utility of Synthetic Chest Radiographs

Abstract: We introduce CheXGenBench, a rigorous and multifaceted evaluation framework for synthetic chest radiograph generation that simultaneously assesses fidelity, privacy risks, and clinical utility across state-of-the-art text-to-image generative models. Despite rapid advancements in generative AI for real-world imagery, medical domain evaluations have been hindered by methodological inconsistencies, outdated architectural comparisons, and disconnected assessment criteria that rarely address the practical clinical value of synthetic samples. CheXGenBench overcomes these limitations through standardised data partitioning and a unified evaluation protocol comprising over 20 quantitative metrics that systematically analyse generation quality, potential privacy vulnerabilities, and downstream clinical applicability across 11 leading text-to-image architectures. Our results reveal critical inefficiencies in the existing evaluation protocols, particularly in assessing generative fidelity, leading to inconsistent and uninformative comparisons. Our framework establishes a standardised benchmark for the medical AI community, enabling objective and reproducible comparisons while facilitating seamless integration of both existing and future generative models. Additionally, we release a high-quality, synthetic dataset, SynthCheX-75K, comprising 75K radiographs generated by the top-performing model (Sana 0.6B) in our benchmark to support further research in this critical domain. We open-source our comprehensive evaluation pipeline, benchmark along with our fine-tuned models, which establish new state-of-the-art performance in synthetic radiograph generation.

References

1. Xie, Enze, et al. "Sana: Efficient high-resolution image synthesis with linear diffusion transformers." arXiv preprint arXiv:2410.10629 (2024).

2. Lin, Tianwei, et al. "HealthGPT: A Medical Large Vision-Language Model for Unifying Comprehension and Generation via Heterogeneous Knowledge Adaptation." arXiv preprint arXiv:2502.09838 (2025).