Scholarlink

To address the challenge of scarce anomalous data and insufficient annotations in medical imaging, this study proposes an Anatomic-Aware Synthetic Lesion CLIP (AASL-CLIP) framework. This method overcomes the limitations of traditional lesion synthesis approaches regarding spatial location and morphology by generating synthetic lesions constrained within lung parenchyma regions of normal chest X-rays using the proposed AnatPaste algorithm; concurrently, an adapter module is designed to accomplish feature space transformation from the natural image domain to the chest X-ray domain. Experiments on the RSNA and CheXpert datasets demonstrate that this method achieves 83.45% and 83.96% AUC on image-level detection tasks, surpassing the best baseline, MediCLIP, by 1.76% and 4.02%, respectively; the generated pixel-level heat maps exhibit high spatial correspondence with lesion regions; the synergistic effect of the AnatPaste algorithm and the adapter module significantly enhances model performance. This framework enables high-precision chest X-ray screening without the need for real anomaly annotations, providing a novel solution for unsupervised anomaly detection in chest X-rays.​

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

Domain-Adapted CLIP with Anatomy-Aware Synthetic Lesions for Unsupervised Chest X-Ray Anomaly Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chengzhi Gui,
  • Meng Xing,
  • Wenjian Liu,
  • Huixin Ma,
  • Hongyu Yang,
  • Wenjing Gong,
  • Yunhui Qu

摘要

To address the challenge of scarce anomalous data and insufficient annotations in medical imaging, this study proposes an Anatomic-Aware Synthetic Lesion CLIP (AASL-CLIP) framework. This method overcomes the limitations of traditional lesion synthesis approaches regarding spatial location and morphology by generating synthetic lesions constrained within lung parenchyma regions of normal chest X-rays using the proposed AnatPaste algorithm; concurrently, an adapter module is designed to accomplish feature space transformation from the natural image domain to the chest X-ray domain. Experiments on the RSNA and CheXpert datasets demonstrate that this method achieves 83.45% and 83.96% AUC on image-level detection tasks, surpassing the best baseline, MediCLIP, by 1.76% and 4.02%, respectively; the generated pixel-level heat maps exhibit high spatial correspondence with lesion regions; the synergistic effect of the AnatPaste algorithm and the adapter module significantly enhances model performance. This framework enables high-precision chest X-ray screening without the need for real anomaly annotations, providing a novel solution for unsupervised anomaly detection in chest X-rays.​