Scholarlink

To address the limitations of traditional diffusion models in generating high-quality, diverse, and realistic medical images, we propose the XR-Diffusion model. This advanced latent diffusion model integrates the Med-BERT text encoder with a visual-language pretraining (VLP) filtering strategy, including classification and image-text retrieval filters, ensuring precise text alignment. By reducing hallucinations, the dual-filtering mechanism significantly enhances image fidelity, diversity, and clinical reliability. XR-Diffusion is ideal for radiology education, AI model training with limited annotated data, and clinical decision support. Evaluations on the MIMIC-CXR and CheXpert datasets show the model achieves FID scores of 12.17 and 17.15, and MS-SSIM scores of 0.12 ± 0.07 and 0.18 ± 0.09, demonstrating exceptional performance.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

XR-Diffusion: A Robust Framework for High-Quality Chest X-ray Synthesis

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jialin Liu,
  • Liping Sun,
  • Weijie Du

摘要

To address the limitations of traditional diffusion models in generating high-quality, diverse, and realistic medical images, we propose the XR-Diffusion model. This advanced latent diffusion model integrates the Med-BERT text encoder with a visual-language pretraining (VLP) filtering strategy, including classification and image-text retrieval filters, ensuring precise text alignment. By reducing hallucinations, the dual-filtering mechanism significantly enhances image fidelity, diversity, and clinical reliability. XR-Diffusion is ideal for radiology education, AI model training with limited annotated data, and clinical decision support. Evaluations on the MIMIC-CXR and CheXpert datasets show the model achieves FID scores of 12.17 and 17.15, and MS-SSIM scores of 0.12 ± 0.07 and 0.18 ± 0.09, demonstrating exceptional performance.