Scholarlink

<p>Current unsupervised anomaly detection methods in chest radiography face challenges in extracting compact spatial features while preserving detailed information, often leading to incomplete recovery of fine-grained image features during reconstruction. To address this, we propose the Radiological Structure Information Extractor (RSIE), a framework comprising three core modules: the parallel attention module (PAM), which enhances feature representation through spatial and channel dimension interactions; the visual enhancement module, which improves visual patterns by optimizing feature selection and integration; and the dual memory matrix (DMM), which combines rich and potential features to maintain both local details and global information during image reconstruction, thus preventing detail loss. Extensive experiments on multiple datasets demonstrate that RSIE significantly improves anomaly detection and image reconstruction performance, validating its superiority over existing methods. The relate code and datasets will be posted at <a href="https://github.com/yigaoyuren/RSIE">https://github.com/yigaoyuren/RSIE</a>.</p>

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

Compact structural feature enhancement for unsupervised anomaly detection in chest radiographs

  • 期刊论文

  • 发表日期：

    2025年8月11日

查看全文
  • Jixun Ye,
  • Wanhui Gao,
  • Yun Wu,
  • Ge Jiao

摘要

Current unsupervised anomaly detection methods in chest radiography face challenges in extracting compact spatial features while preserving detailed information, often leading to incomplete recovery of fine-grained image features during reconstruction. To address this, we propose the Radiological Structure Information Extractor (RSIE), a framework comprising three core modules: the parallel attention module (PAM), which enhances feature representation through spatial and channel dimension interactions; the visual enhancement module, which improves visual patterns by optimizing feature selection and integration; and the dual memory matrix (DMM), which combines rich and potential features to maintain both local details and global information during image reconstruction, thus preventing detail loss. Extensive experiments on multiple datasets demonstrate that RSIE significantly improves anomaly detection and image reconstruction performance, validating its superiority over existing methods. The relate code and datasets will be posted at https://github.com/yigaoyuren/RSIE.